Module 1 — Appointment Booking & Management

A centralized, real-time appointment management system for walk-in and online bookings:  Schedule, reschedule, and cancel appointments with a single click — centralized dashboard with real-time queue and wait times  Support for walk-in and online appointment channels — quick-access buttons for both  Patient registration at booking — captures name, gender, DOB, mobile, Aadhaar/ABHA ID  New ABHA account creation directly from the appointment interface — or verification of existing ABHA ID  Real-time appointment tracking — arrival status, wait time, consultation status all visible in one view  View patient vitals, medical history, prescriptions, and previous examination notes directly from the appointment dashboard  Action controls: view detail, update time, reschedule, cancel — with confirmation prompts  Department and doctor selection — appointments routed to the correct consultant automatically  Integration with ABHA for secure digital health record access and sharing at the point of booking

 

Module 2 — Patient Management

Complete end-to-end patient record management — from first registration to ongoing care:  Unique UHID (Universal Health ID) generated for every patient at registration  Comprehensive patient profile: demographics, medical history, allergy triggers, vital alerts, immunisation records  Visit logs and doctor assignment history across all OPD and IPD interactions  Consent logs and documented approvals — legally compliant patient consent management  Integrated view across prescriptions, appointments, billing, and lab reports — all accessible from a single patient record  Vital signs dashboard: height, weight, BMI, temperature, pulse, blood pressure — all logged per visit  ABHA / Aadhaar linking — patient records linked to national health identity for interoperability

 

Module 3 — Laboratory Management (Diagnostic Centre)

End-to-end lab workflow management — from test assignment to report delivery — purpose-built for diagnostic centres:  Test assignment from doctor to lab — linked directly to patient profile by name, ID, or ABHA number  Sample ID generation and tracking — collection date, dispatch, acknowledgement, processing status all tracked  Priority-based sample handling — urgent samples flagged and processed with higher priority  Result entry: both automated (machine connectivity) and manual input options  Automated report verification alerts — notifies lab doctors and physicians when results are ready for review  PDF lab report generation with one click — auto-distributed to patient, treating doctor, and relevant staff  Outsourced sample management — supports samples sent to external reference labs  Integrated patient history — view all past tests and results from the same interface  ABHA-compatible — search and link patients using national health ID for paperless lab processing  Role-based access — lab assistants, technicians, and consultants each see only relevant data and actions  Diagnostic machine connectivity — direct integration for automated result capture, reducing manual errors

 

Module 4 — AI Prescription Generation & Clinical Intelligence

The AI layer is the most powerful differentiator in Ektos Health — enabling doctors to generate accurate, formatted prescriptions in under a minute using voice or text, with AI-assisted diagnosis support:  Voice-to-Text Prescription Dictation  Doctors dictate the prescription verbally — the system converts speech to structured prescription text in real time  Supports medical terminology, drug names, dosages, and clinical instructions accurately  Significantly reduces consultation time — doctors spend less time typing and more time with the patient  Dictated text is editable before finalisation — doctor retains full clinical control  AI-Generated Diagnosis Suggestions  Doctor enters the patient’s chief complaint and clinical notes — AI analyses these against the patient’s previous diagnosis history  AI suggests probable diagnoses ranked by likelihood — based on the symptom pattern, history, and clinical context  Differential diagnosis support — highlights conditions to rule out based on the presenting complaint  AI draws from the patient’s full longitudinal history stored in the system — not just the current visit  Doctor reviews, selects, or overrides AI suggestions — the system learns and improves with usage  AI Prescription PDF Generation  Once the consultation is complete, the system generates a professionally formatted, branded prescription PDF automatically  Prescription includes: patient details, date, doctor name and registration number, diagnosis, medications with dosage and frequency, instructions, follow-up date  Compliant with ABDM digital prescription standards — can be shared electronically via ABHA  Prescriptions stored against the patient record — visible in appointment history and patient profile  Configurable prescription templates per department or doctor — managed via Settings module

 

Module 5 — Admission & Discharge (IPD)

Full inpatient lifecycle management — from bed allocation on admission to discharge summary generation:  Digital patient registration at admission — captures insurance details, ID, health history, and emergency contacts  Intelligent bed allocation — auto-assigns beds based on availability, room type (ICU, General, Private, Isolation), and clinical condition  Real-time bed management dashboard — live view of all bed IDs, types, room numbers, and occupancy status  Bed category management — ICU, General, Private, Isolation configured and managed via the system  Real-time patient condition monitoring — clinicians and nursing staff track status updates and discharge readiness  Discharge planning — clinicians set discharge timelines and manage pending reports before release  Automated discharge summary generation — produced with a few clicks, including diagnosis, treatment summary, medications, and follow-up instructions  Bed released automatically on discharge — status updated from ‘In Use’ to ‘Available’ in real time  Cross-department coordination — admission, billing, nursing, and housekeeping all notified of status changes  Care team collaboration dashboard — doctors, nurses, and admin work from the same centralized view

 

Module 6 — Billing Management

Integrated, transparent billing across all hospital services — from OPD consultations to lab tests and inpatient stays:  Automated bill generation — charges for consultations, lab tests, procedures, medications, and room rental all captured automatically  Multi-status billing — tracks paid, unpaid, partially paid, and refunded invoices  Detailed payment breakdown per patient — line-item visibility of every charge  Patient-wise billing history — all bills and payment records linked to the patient profile  Search and filter tools — quickly locate invoices by patient, date, doctor, or billing status  Consent-based billing — billing actions linked to documented patient consent  Invoice and billing templates configurable per hospital — managed via Settings module

 

Module 7 — Settings, Master Data & System Configuration

Full control over how Ektos Health operates within your institution:  Brand settings — hospital logo, colours, and name applied across all reports, prescriptions, and invoices  Email and SMS configuration — automated notifications for appointment confirmations, lab results, and billing  Invoice, prescription, appointment, and lab report templates — all configurable without developer involvement  User roles and access control — granular permissions per role (doctor, nurse, lab technician, billing staff, admin)  Compliance and audit trails — all actions logged with user ID, timestamp, and action type  Department management — add, edit, and deactivate departments and specialties  Doctor and staff profiles — registration numbers, specialties, and schedule management  Billing item setup — configure service codes, pricing, and insurance billing items  Real-time system updates — configuration changes take effect immediately without downtime

ABDM Compliance	Full integration with Ayushman Bharat Digital Mission — ABHA creation, verification, and health record linkage built into appointment, lab, and prescription workflows
ABHA Integration	Patients can create a new ABHA (Ayushman Bharat Health Account) or link an existing one at registration — enabling secure nationwide health record interoperability
Aadhaar Verification	Aadhaar-based patient identity verification supported at registration and appointment booking
Digital Prescriptions	AI-generated prescription PDFs comply with ABDM digital prescription standards — shareable via ABHA
NABH Standards	Workflow design, documentation requirements, consent management, and audit trails align with NABH accreditation criteria
Audit Trails	All clinical and administrative actions logged with user, timestamp, and action type — supporting NABH audit requirements
Data Security	Patient health data encrypted at rest and in transit — role-based access prevents unauthorised data exposure
Compliance Readiness	90% compliance readiness score across ABDM/NABH requirements — hospitals go live already meeting accreditation benchmarks

Technology & Architecture

Ektos Health is a cloud-native application built on modern web technologies, deployed on scalable cloud infrastructure:

Frontend	React.js — fast, responsive, mobile-compatible web interface — works on desktop, tablet, and mobile without a separate app
Backend	Node.js / Python REST APIs — microservice architecture for each module (appointments, lab, billing, AI)
Cloud	Cloud-hosted — scalable, high-availability infrastructure with 24/7 uptime monitoring
Database	Secure relational database with full patient record history, audit logs, and real-time updates
AI Engine	GPT-4 integration for diagnosis suggestions and prescription generation — custom prompt engineering for clinical context
Voice-to-Text	Speech recognition API with medical vocabulary training — dictation converted to structured prescription text
PDF Engine	Automated prescription and lab report PDF generation — templated, branded, and configurable per department
ABDM API	Official ABDM API integration — ABHA creation, verification, and health record exchange
Notifications	SMS and email notification engine — appointment reminders, lab result alerts, billing confirmations
Security	Role-based access control, encrypted data transmission, session management, and full audit logging

AI Features — How They Work

AI Feature	How It Works	Clinical Benefit
Voice-to-Text Prescription	Doctor speaks the prescription aloud. Speech recognition converts dictation to structured text — drug names, dosages, instructions — in real time. Doctor reviews and confirms before saving.	Consultation time reduced significantly. Doctor maintains eye contact with patient rather than typing. Reduces transcription errors.
AI Diagnosis Suggestion	Doctor enters chief complaint and clinical notes. AI analyses these against the patient’s stored diagnosis history and symptom patterns to suggest probable diagnoses ranked by likelihood.	Supports junior doctors with differential diagnosis. Reduces diagnostic oversight. Highlights conditions to rule out. Improves clinical decision quality.
AI Prescription PDF	Once diagnosis and treatment are confirmed, the system auto-generates a formatted, branded prescription PDF including patient details, diagnosis, medications with dosage and frequency, and follow-up date.	Professional, legible, legally compliant prescriptions every time. Stored against patient record. Shareable via ABHA. No formatting overhead for doctor.
Previous Diagnosis Context	AI accesses the patient’s full longitudinal history — all previous diagnoses, medications, allergies, and test results — as context when generating suggestions for the current visit.	Prevents prescribing contraindicated medications. Highlights recurring conditions. Improves continuity of care across multiple visits and doctors.

Implementation Approach

Ektos Health is deployed as a cloud SaaS product. New hospitals and diagnostic centres can be onboarded and live within days:

Day 1–2 — Setup	Hospital profile, branding, departments, doctors, and billing items configured in Master Data and Settings. User roles and access levels assigned.
Day 2–3 — Integration	ABDM/ABHA API connection configured. SMS/email notification templates set up. Diagnostic machine connectivity tested where applicable.
Day 3–4 — Data Entry	Existing patient records migrated or entered. Lab test catalogue configured. Room and bed categories set up for IPD centres.
Day 4–5 — Training	Staff training on appointment, lab, billing, and prescription workflows. Doctors trained on voice-to-text and AI diagnosis features.
Day 5 — Go-Live	Live patient appointments and consultations begin. Peritos Solutions provides hypercare support for the first two weeks post-launch.
Ongoing	24/7 cloud support. Feature updates deployed automatically. AI model improves with usage. ABDM regulatory updates applied centrally.

Client Testimonials

“Ektos Health has significantly streamlined our operations and improved efficiency across our healthcare workflows. The platform is intuitive, reliable, and backed by a highly responsive team at Peritos Solutions.”  — Akanksha Niranjan, Director, Ekanshi Solutions, Lucknow

“The platform has simplified our day-to-day clinic management, from patient records to reporting. It has reduced manual work and improved overall efficiency. The support team is proactive and always available when needed.”  — Dr. Ramnath Mishra, Clinic, Bhubaneswar

“Ektos Health has transformed how we manage patient records and hospital operations. The system is easy to use, secure, and provides quick access to critical information, enabling us to deliver better patient care.”  — Dr. Anil Chowdhary, Lal Kothi Hospital, Jaipur

About the Client

Bayleys Real Estate is New Zealand’s largest full-service real estate agency, operating across residential, commercial, rural, and property management services across New Zealand and Australia. With over 150 property management agents generating 1,800 to 2,000 rental appraisals per month, the manual documentation workload was significant, inconsistent across regions, and a major drain on agent productivity.

The leadership team — including the National Director and Financial Director — identified AI-driven automation as the strategic priority, with a clear mandate: cut appraisal time, improve consistency and accuracy, and free agents to focus on client relationships rather than documentation.

Project Background

Bayleys was already running a modern AWS-hosted property search web application built by Peritos Solutions — giving agents a cloud-based way to search and manage property records from any device. The next phase was to add AI intelligence on top of that foundation.

Each rental appraisal required an agent to manually research comparable properties, write a property description, calculate a market rent range, pull together agent branding and contact details, and format everything into a professional PDF. This 30 to 45 minute process, repeated thousands of times per month, was the single biggest time drain in the property management workflow.

Peritos Solutions was engaged to design and build the AI upgrade: automating the appraisal document, training a mathematical valuation model on historical data, deploying an AI copywriting engine for property descriptions, and building the AskKen AI chatbot for on-demand knowledge access.

Requirements

• Generate a complete, professional, branded rental appraisal report in under one minute — from a single property address entry
• AI-generated property description — trained on thousands of historical Bayleys appraisals to match the tone and style of an experienced property manager
• EMV (Estimated Market Rent) calculated by a machine learning model trained on historical sales and rental data — accuracy within ±3% of experienced agent assessment
• Property attributes (bedrooms, bathrooms, floor area, carparks) automatically retrieved via CoreLogic API
• Property listing image automatically pulled via Bayleys API where the property is currently listed for sale
• Agent profile photo, name, contact details, and office auto-populated via Office 365 Single Sign-On
• School zones and local amenities automatically identified from property address
• Support for multiple property types — house, apartment, unit, minor dwelling, home and income
• Manual entry flow for new-build and off-plan properties not yet in CoreLogic or Bayleys systems
• AskKen AI chatbot — zero-training, conversational interface covering legislation, tribunal cases, market data, maintenance costs, yield calculations, and suburb intelligence
• RAG architecture grounding all AI responses in controlled Bayleys proprietary data — not the open internet

Scope & Feature List

Module 1 — Automated Rental Appraisal Report

A property manager enters or confirms the property address. The system generates a complete, professionally formatted, Bayleys-branded rental appraisal report automatically:  Property address triggers automatic data retrieval from CoreLogic API — bedrooms, bathrooms, floor area, carparks, property type  Property image pulled from Bayleys listing API if the property is currently listed for sale — ensuring brand-consistent presentation  Agent profile, photo, and contact details auto-populated via Office 365 SSO — the system knows who is logged in  EMV rental range calculated and displayed — powered by the Random Forest valuation model (see Module 2)  AI-generated property description inserted automatically — see Module 3  Rental details section populated: current market value range, property type, report date Back-page advert customisable by region or individual agent — uploaded via application settings  Multiple property types supported — houses, apartments, units, minor dwellings, home and income  Manual entry available for new-build and off-plan properties — feeds directly into the Bayleys Data Lake  Output: professional PDF matching Bayleys national branding guidelines — consistent across all 150+ agents and all regions

Solution Architecture

Technology & Architecture

The AI platform is built cloud-native on AWS, with OpenAI -GPT-4O powering the language generation layer. All AI responses are grounded in controlled proprietary data — not the open internet.

Layer	Technology / Service	Role
Cloud	AWS (primary)	Serverless infrastructure, Lambda functions, API Gateway, DynamoDB, S3, SNS, CloudWatch
AI / LLM	OpenAI GPT-4O	Base generative and reasoning capability for AskKen AI and property description generation
AI Orchestration	AWS Lambda + Node.js/Python	Microservice orchestrating: user input → RAG retrieval → LLM call → answer filtering → UI response
EMV Model	Random Forest Regression	Trained on Bayleys Data Lake — bootstrap aggregation, geometric mean, Pearson R suburb weighting — ±3% accuracy
RAG Layer	Vector store + proprietary docs	54,000 tribunal cases + legislation + market data indexed — retrieved at query time to ground LLM responses
Data Layer	Bayleys Data Lake	Historical appraisals, rental data, property records, new-build manual entries — feeds EMV model training
Property Data	CoreLogic API	Bedrooms, bathrooms, floor area, carparks — retrieved automatically on address entry
Images	Bayleys Listings API	Current listing photos pulled into appraisal report automatically for listed properties
Auth	Office 365 SSO	Single sign-on — agent profile, photo, and contact details auto-populated in every report
Security	AWS WAF + token auth	Web Application Firewall + token-based API endpoints — all property and AI data secured
Guardrails	Rules engine	LLM outputs reviewed against compliance checklists; unsupported assertions stripped before reaching the user

Application Images

Implementation Approach

The project kicked off in October 2024 and delivered an MVP into UAT for Auckland agents by early March 2025 — on schedule and within budget:

Phase 1 — Discovery	Requirements workshops, CoreLogic and Bayleys API integration scoping, data lake assessment, RAG document inventory, architecture design on AWS
Phase 2 — EMV Model	Initial Random Forest baseline, geometric mean refinement, Pearson R suburb-level correlation weighting — iterated until ±3% accuracy achieved nationally
Phase 3 — Appraisal Engine	Automated report generation, CoreLogic integration, Bayleys API image pull, Office 365 SSO, multiple property type handling, 90-day expiry logic
Phase 4 — AI Descriptions	GPT-4O fine-tuning on historical Bayleys appraisals, sentiment analysis training, school zone and amenity integration, review and editing workflow
Phase 5 — AskKen AI	RAG pipeline build, document ingestion (legislation + 54,000 tribunal cases), guardrails layer, GPT-4O prompt engineering, chat interface on mobile and desktop
Phase 6 — UAT & Go-Live	UAT with Auckland property managers, regional disclaimer configuration, performance tuning, cost optimisation on AWS, go-live and hypercare

Challenges & Solutions

 

EMV accuracy from ±15% to ±3%	Initial models had very high variance. Three refinement stages — basic features, geometric mean aggregation, then Pearson R local correlation weighting — drove accuracy to ±3% nationwide, matching experienced agent assessments.
AI descriptions that sound human	Generic AI property descriptions were immediately identifiable and not fit for purpose. Peritos trained the model on thousands of real historical Bayleys appraisals using sentiment analysis — the output now matches Bayleys‘ own writing style.
New-build data gaps	Off-plan and new-build properties lack CoreLogic data. A manual entry flow was built, with all entered data feeding back into the Bayleys Data Lake to improve future EMV accuracy in newly developed areas.
RAG grounding vs. hallucination	LLMs are prone to confident but incorrect legal answers. All AskKen AI responses are grounded in indexed proprietary documents with a guardrails layer that strips unsupported assertions before they reach the agent.
54,000 tribunal case ingestion	Ingesting and indexing this volume of case law required careful document parsing, metadata tagging by jurisdiction and date, and chunking strategy to ensure relevant cases are retrieved at query time.
First AI project — high visibility	As Bayleys‘ inaugural AI initiative this project set the benchmark for all future AI investment. Peritos delivered on time, within budget, with quantifiable ROI — and secured a $20,000 Microsoft contribution recognising it as an industry first.

Financial Impact

Appraisal time reduction	From 30–45 minutes per appraisal to under 1 minute — saving approximately 40 minutes per appraisal
Monthly appraisal volume	1,800–2,000 appraisals per month (60,000–67,000 annually)
Saving per appraisal	NZD $33.33 per appraisal (at NZD $50/hr agent cost)
Annual appraisal saving	NZD $720,000–$800,000 per year from appraisal automation alone
AskKen AI saving	10 hrs manual research saved per agent/month × 150 agents × NZD $50/hr = NZD $75,000/month
Annual AskKen saving	NZD $900,000+ per year in research, compliance checking, and legal advisory time
Total annual savings	NZD ~$1.75 million per year — combined appraisal automation + AskKen AI

Key Benefits

• Appraisal turnaround time reduced from 30–45 minutes to under 1 minute — freeing agents to focus on client relationships
• AI-generated property descriptions match the tone and style of experienced Bayleys property managers — consistent quality across all agents and regions
• EMV accuracy of ±3% nationwide — matching or exceeding the estimates of experienced agents
• AskKen AI gives every property manager instant access to legislation, tribunal cases, market data, and maintenance knowledge — without leaving the platform
• Agent profile, contact details, and property images auto-populated via SSO and API integrations — zero manual formatting
• New-build and off-plan data captured manually feeds back into the Bayleys Data Lake — building proprietary market intelligence ahead of public availability
• Stress and cognitive load reduced for property managers — repetitive tasks automated, accuracy protected
• Platform scales across New Zealand and Australia from a single codebase on AWS
• Recognised by Microsoft with a $20,000 contribution as an industry-first AI innovation

Support & Next Steps

Peritos Solutions provided post-go-live hypercare covering AI output quality monitoring, EMV model tuning, and integration stability. Automated pipelines re-index the RAG knowledge base as new tribunal decisions and legislation changes arrive — the system improves continuously without manual intervention.

Planned next phase:

• Extension to Australian markets — McGrath Real Estate agent base
• Expanded AskKen AI modules covering insurance policy guidance and advanced financial modelling
• Business intelligence dashboard — appraisal volumes, lead conversion rates, and sales-to-property-management referral tracking
• Further EMV model refinement with expanded suburb-level correlation data from new-build entries
• Automated contact creation in the platform when new landlords are onboarded

About the Client

Wine-Searcher is the world’s most visited wine marketplace and price comparison platform, connecting consumers, collectors, and trade buyers with wine retailers across more than 180 countries. With over 8 million distinct wine listings and millions of monthly active users, Wine-Searcher depends on accurate, consistent, and richly detailed product data to power search, recommendations, and pricing intelligence.

The editorial and data team had long relied on a combination of contributor-submitted data and manual verification — a process that created bottlenecks as the catalogue scaled and left the majority of listings without historical or regional context. The leadership team identified AI-driven automation as the path to consistent data quality at scale, with a clear brief: recognise the label, extract the facts, and tell the story.

Project Background

Wine-Searcher already operated on AWS — with a modern data pipeline handling listing ingestion, search indexing, and pricing data. The next phase was to add AI intelligence on top of that foundation: automatically reading wine labels, extracting structured metadata, and generating contextualised provenance summaries that elevate the consumer experience.

Each label presents a unique challenge: typography varies wildly across producers and countries, regulatory text overlaps with brand content, multi-language labels require translation before parsing, and the historical context for a given wine may span decades of regional winemaking history. No off-the-shelf model could handle the full pipeline — a purpose-built solution on SageMaker was required.

Peritos Solutions was engaged to architect and build the end-to-end AI pipeline: image ingestion, a custom label recognition model, OCR-based structured extraction, a RAG-grounded Bedrock summarisation layer, and integration back into Wine-Searcher’s live listings platform.

Requirements

• Automatically identify winery, wine name, vintage, appellation, grape variety, alcohol percentage, and certifications from a label image — accuracy target of >92%
• OCR extraction of all label text, structured into a validated JSON schema compatible with Wine-Searcher’s existing listing data model
• AI-generated historical provenance summary — 150 to 250 words, grounded in a curated wine knowledge base covering regions, producers, vintages, and critics’ scores
• Support for labels in French, Italian, German, Spanish, Portuguese, and English — with automated language detection and translation before structured parsing
• Inference latency under two seconds per label for real-time API use; batch enrichment pipeline for retroactive processing of existing catalogue
• Confidence scoring on every extracted field — low-confidence predictions routed to a human review queue rather than published directly
• RAG grounding for Bedrock summaries — no hallucinated facts about producers, vintages, or regions; all claims traceable to indexed knowledge base sources
• Continuous model improvement — validated human corrections feed back into the SageMaker feature store and trigger incremental retraining
• SageMaker Model Monitor integration — automated data drift detection and accuracy degradation alerts with retraining triggers
• AWS-native security — Cognito authentication, WAF rate limiting, end-to-end encryption, no label image data retained beyond processing window

Scope & Feature List

Module 1 — Label Recognition Engine (Amazon Rekognition + SageMaker)

A wine label image is submitted via API or batch upload. A custom Convolutional Neural Network (CNN) model — trained on a corpus of 500,000+ labelled wine label images and hosted on a SageMaker real-time endpoint — identifies the winery, wine name, appellation, grape variety, vintage year, and alcohol content from the visual label. The model handles skewed angles, low-resolution mobile uploads, partial label occlusion, and multi-label bottle formats (front + back). A confidence score is assigned to each detected field. Fields below the confidence threshold are flagged for the human-in-the-loop review queue rather than written directly to the listing. The SageMaker endpoint is versioned — A/B testing allows new model versions to serve a percentage of traffic before full promotion, with automatic rollback if accuracy degrades. Model training used SageMaker Automatic Model Tuning to optimise hyperparameters across CNN architecture variants.

Module 2 — OCR Structured Extraction (Amazon Textract)

Following visual label recognition, Amazon Textract performs OCR on the full label image — extracting all printed text with bounding box coordinates and confidence scores. A post-processing Lambda function, running in Node.js, applies a Wine-Searcher–specific parsing schema to the raw Textract output: regulatory text is separated from brand content; dates are resolved using regional wine labelling conventions; certifications (organic, biodynamic, AOC, DOC, DOCG) are identified by keyword matching against a controlled vocabulary; and alcohol percentages are extracted from the mandatory legal text band. Multi-language labels are detected by AWS Comprehend and routed through a translation Lambda before parsing. The structured output is validated against the Wine-Searcher listing schema before being written to DynamoDB and pushed to the listings enrichment queue.

Module 3 — Historical Provenance Summary (Amazon Bedrock + RAG)

Once structured label data is confirmed, a RAG pipeline generates a 150 to 250 word historical provenance summary for each wine. Proprietary knowledge base documents — covering wine regions, appellations, major châteaux and producers, vintage quality guides, and critical scoring context — are ingested, chunked, and indexed in an Amazon OpenSearch vector store. At query time, the structured label data (winery, appellation, vintage, grape variety) is used to retrieve the most relevant knowledge base passages. These passages, together with the structured label metadata, are passed to Amazon Bedrock (Claude) as context. The model generates a provenance summary grounded entirely in the retrieved documents — hallucination guardrails strip any claim not traceable to an indexed source. Example output: ‘Château Margaux 2015 is a First Growth Bordeaux from the Margaux appellation of the Médoc. The 2015 vintage was widely acclaimed as one of the finest of the decade — warm, dry conditions produced exceptional concentration, with the Wine Advocate awarding 100 points. Produced primarily from Cabernet Sauvignon with Merlot, Petit Verdot, and Cabernet Franc, this wine is expected to peak between 2030 and 2060.’

Module 4 — Human-in-the-Loop Review & Continuous Learning

Low-confidence label predictions and any Bedrock summary flagged by the guardrails layer are routed to a Wine-Searcher editorial review queue — a lightweight web UI where specialists can confirm, correct, or reject AI outputs before publication. All validated corrections are written back to the SageMaker feature store. An automated retraining trigger fires when the volume of corrections for a given producer category exceeds a configurable threshold — a SageMaker Pipeline runs the incremental training job on the updated feature set, evaluates accuracy against a holdout set, and promotes the new model version to the endpoint if performance improves. This closed loop ensures the model improves continuously as Wine-Searcher’s catalogue grows and new producers are encountered.

Solution Architecture

The platform is built cloud-native on AWS. All AI inference runs within the Wine-Searcher AWS account — no label image data leaves the AWS environment. The architecture is fully serverless outside of the SageMaker inference endpoint, with pay-per-request Lambda functions handling orchestration, parsing, validation, and API integration.

AWS architecture — API Gateway → Lambda orchestration → SageMaker endpoint (label recognition) → Textract OCR → Bedrock RAG summarisation → DynamoDB enrichment store → listings API push → CloudWatch + Model Monitor

Technology & Architecture

Layer	Technology / Service	Role
Cloud	AWS (primary)	Serverless infrastructure — Lambda, API Gateway, DynamoDB, S3, SNS, CloudWatch
Label Recognition	Amazon Rekognition + SageMaker	Custom CNN model trained on wine label images — winery, vintage, appellation, grape variety detection
OCR Extraction	Amazon Textract	Extracts structured text from label scans — wine name, producer, region, alcohol %, certifications
AI Summaries	Amazon Bedrock (Claude)	Generates historical provenance summaries from extracted structured data + retrieved knowledge base context
ML Training	Amazon SageMaker	Model training, versioning, A/B endpoint testing, feature store, and batch inference pipeline
RAG Layer	SageMaker + OpenSearch	Wine knowledge base (regions, châteaux, vintages, critics) indexed and retrieved at query time to ground Bedrock summaries
Data Pipeline	AWS Glue + S3 Data Lake	Label image ingestion, transformation, enrichment, and storage — feeds training pipeline and inference cache
Listings API	Wine-Searcher Internal API	Pushes enriched label metadata and AI summaries back into live product listings
Auth & Security	AWS Cognito + WAF	Secure API access — rate limiting, token auth, IP-based WAF rules protecting the inference endpoint
Monitoring	CloudWatch + SageMaker Model Monitor	Data drift detection, prediction quality tracking, automated retraining triggers when accuracy degrades

Implementation Approach

Phase 1 — Discovery	Requirements workshops, label image corpus assessment, Data Lake scoping, AWS architecture design, SageMaker feasibility study, RAG knowledge base inventory
Phase 2 — Data Pipeline	Label image ingestion pipeline via AWS Glue, S3 staging and normalisation, Textract OCR extraction, structured JSON output schema definition
Phase 3 — SageMaker Model	CNN training on labelled wine image dataset, SageMaker hyperparameter tuning, A/B endpoint deployment, accuracy benchmarking against manual expert classification
Phase 4 — RAG & Bedrock	Wine knowledge base build (regions, châteaux, vintages, critics), OpenSearch indexing, Bedrock (Claude) prompt engineering, hallucination guardrails, summary quality evaluation
Phase 5 — API & Listings Integration	Wine-Searcher internal API integration, enriched metadata push back to listings, search index update pipeline, caching layer for inference results
Phase 6 — UAT & Go-Live	Accuracy UAT with Wine-Searcher editorial team, performance and cost tuning on AWS, SageMaker Model Monitor activation, go-live across 8M+ listings, hypercare

Challenges & Solutions

Low label image quality	Many label images in the marketplace were low-resolution, tilted, or partially obscured. SageMaker training used augmentation pipelines (rotation, blur, contrast variation) to make the model robust across real-world upload quality.
Multi-language label text	Wine labels appear in French, Italian, German, Spanish, Portuguese, and English. Textract was supplemented with a language-detection Lambda routing non-English extractions through a translation layer before structured parsing.
Vintage ambiguity	Older bottles often carry multiple dates (bottling, release, vintage). A custom post-processing rule engine was built to resolve date ambiguity using regional norms (e.g. Bordeaux vs. New World labelling conventions).
Hallucination in historical summaries	Early Bedrock outputs included plausible-sounding but fabricated château histories. A RAG grounding layer with a curated wine knowledge base (regions, producers, vintages, critics’ scores) was introduced — all summaries now cite indexed sources.
Cold-start for rare producers	Small-production wineries had few training images. A human-in-the-loop review queue was built so that low-confidence predictions are flagged for expert validation and fed back into the SageMaker feature store for incremental retraining.
Scale of batch enrichment	Retroactively enriching 8M+ existing listings required a SageMaker batch transform job spread across spot instances — completing in 72 hours at a fraction of on-demand cost, with checkpointing to handle interruptions gracefully.

Financial Impact

Time saved per label	15–20 minutes manual research and transcription reduced to under 2 seconds automated
Monthly new label volume	60,000+ new SKUs ingested monthly across contributor uploads and editorial additions
Saving per label	USD $12–$16 per label at USD $50/hr data team cost (conservative estimate)
Annual data team saving	USD $8.6M–$11.5M per year from automated label enrichment alone
Historical catalogue enrichment	8M+ existing listings retroactively enriched — equivalent of 1,600+ person-years of manual work completed in 72 hours via batch transform
Consumer engagement uplift	Listings with provenance summaries show 34% higher click-through and 22% higher add-to-cart rates vs unenriched listings
Total annual value	USD $10M+ combined — data team savings + engagement-driven revenue uplift

Key Benefits

• Label recognition and structured data extraction in under 2 seconds — replacing 15 to 20 minutes of manual research per SKU
• Provenance summaries grounded in a curated wine knowledge base — no hallucinated producer histories, no fabricated critical scores
• Multi-language support across French, Italian, German, Spanish, Portuguese, and English — single pipeline, zero separate tooling
• Confidence-scored outputs with a human-in-the-loop queue — accuracy protected, with validated corrections feeding continuous model improvement
• SageMaker Model Monitor detecting data drift and triggering retraining — the model stays accurate as new producers and regions are encountered
• 8M+ legacy listings retroactively enriched via batch transform — years of backlog cleared in 72 hours at spot instance cost
• 34% higher click-through and 22% higher add-to-cart on enriched listings — provenance data demonstrably drives consumer engagement
• Fully AWS-native — no label image data leaves the Wine-Searcher AWS environment; WAF, Cognito, and end-to-end encryption throughout
• Serverless outside of the SageMaker endpoint — scales to any ingestion volume without infrastructure management

Support & Next Steps

Peritos Solutions provided post-go-live hypercare covering model accuracy monitoring, Bedrock summary quality review, and AWS infrastructure optimisation. Automated pipelines re-index the wine knowledge base as new vintage guides, regional legislation, and critic publications are added — the RAG layer stays current without manual intervention.

Planned next phase:

• Sommelier AI — a conversational Bedrock-powered assistant integrated into Wine-Searcher’s consumer app, capable of food pairing, cellar management advice, and vintage comparison
• Fake label detection — a secondary SageMaker model trained to identify counterfeit and mislabelled bottles using visual and metadata anomaly detection
• Auction & secondary market intelligence — provenance summaries extended with auction price history, rarity scoring, and investment potential indexing
• Visual similarity search — a SageMaker embedding model enabling consumers to search by label image rather than text query
• Expansion of the wine knowledge base — additional appellations, small-production natural wine producers, and emerging New World regions

About HR Mind

HR Mind is a global resourcing company founded in 2010 that offers end-to-end recruitment and HR solutions to organisations in domestic and international markets. With deep expertise across multinational and local businesses, HR Mind provides tailored talent acquisition solutions across eight industry verticals: Infrastructure/EPC, Internet/E-Commerce, Renewable Energy, Automotive, FMCG, Information Technology, Healthcare, and Industrial/Manufacturing.

The company serves four distinct client segments — MNCs, SMEs, Startups, and Joint Ventures — placing candidates at senior and middle management levels as well as running global talent acquisition mandates. Beyond recruitment, HR Mind also provides HR solutions including payroll outsourcing, salary analysis, and candidate assessment services.

With a large and growing pipeline of job openings across multiple industries and client types, HR Mind faced a scalability challenge: the volume of resumes received per job opening was growing faster than their recruiter team could manually process. A technology solution was needed to automate the screening step without sacrificing the quality and accuracy that HR Mind’s clients expected.

The Problem

Recruitment at scale is fundamentally a data matching problem — but one that had been solved manually for decades. HR Mind identified several specific pain points driving the need for an AI solution:

• High volume, low signal — hundreds of resumes received per job opening, with the majority not matching the role requirements. Manually reading each one to determine relevance was the single biggest time drain in the recruitment process
• Inconsistent screening — different recruiters applied different criteria when reviewing the same JD, leading to inconsistent shortlists and missed candidates
• JD complexity — Job Descriptions often contain 20–40 specific skills, qualifications, and experience requirements. Matching these manually against a resume was error-prone and incomplete
• Keyword blindness — resumes use different terminology for the same skills (e.g. ‘ML’, ‘Machine Learning’, ‘Artificial Intelligence’, ‘Deep Learning’) — manual reviewers often missed valid candidates due to vocabulary differences
• No objective scoring — shortlisting was subjective. There was no quantitative score to explain why one candidate ranked above another, making it difficult to justify shortlists to clients
• Slow time-to-shortlist — delivering a qualified candidate shortlist to a client took 2–5 days from job opening. In competitive talent markets, this delay cost placements
• Inability to re-use candidate pool — past resumes in the database were not being systematically re-matched against new job openings — a significant lost opportunity

Scope & Feature List

Solution Architecture

The platform is built entirely on AWS serverless architecture — no servers to manage, automatic scaling with application volume, and pay-per-use pricing that keeps costs proportional to usage:

Layer	AWS Service / Technology	Role in HR Mind Platform
Ingestion	AWS S3	Secure storage for all uploaded resumes (PDF, Word, text) — triggers Lambda functions on upload for automatic processing
Document Parse	AWS Textract	Extracts structured text from PDF resumes including scanned documents — feeds clean text to the NLP pipeline
NLP / AI	AWS SageMaker + Lambda	ML models for resume classification, keyword extraction, semantic matching, and candidate scoring — deployed as serverless endpoints
Text Embeddings	Amazon Bedrock	NLP entity extraction and semantic understanding — identifies skills, job titles, organisations, and dates from unstructured resume text
Matching Engine	AWS Lambda	Serverless function orchestrating JD parsing, keyword extraction, resume-to-JD matching, and score calculation — triggered per job opening
API Layer	Lambda Direct URl	RESTful API endpoints for recruiter dashboard, resume upload, JD submission, shortlist retrieval, and candidate search
Data Store	SQL DB	NoSQL database storing structured candidate profiles, match scores, JD keyword profiles, shortlists, and recruiter actions
Auth	Auth0	Recruiter and admin authentication — role-based access to platform features and candidate data
Monitoring	AWS CloudWatch	Performance monitoring, error alerting, Lambda invocation metrics, and cost tracking dashboards
Security	AWS WAF + IAM	Web Application Firewall protecting API endpoints; IAM roles enforcing least-privilege access to all AWS resources
Storage Archive	S3 Lifecycle Policies	Automatic archival of old resumes to S3 Glacier — cost management for long-term candidate data retention

End-to-End Workflow

Step 1 — JD Upload	Recruiter uploads or pastes the Job Description into the platform. The NLP engine parses it and extracts a structured keyword profile with weighted categories.
Step 2 — Keyword Review	Recruiter reviews the extracted keywords, adjusts weights if needed, adds custom terms, and confirms the JD profile. The AI is now primed for matching.
Step 3 — Resume Upload	Resumes uploaded in bulk (PDF/Word) via drag-and-drop or sourced from the existing candidate database. AWS S3 stores each file and triggers automatic processing.
Step 4 — Classification	Each resume is classified into a role category by the ML model. The NLP pipeline extracts structured data: skills, experience, education, job titles, and tenure.
Step 5 — Keyword Matching	Each parsed resume is matched against the JD keyword profile. Exact and semantic matches are scored. Skills gaps are identified. An overall match score (0–100) is calculated.
Step 6 — Review & Export	Recruiter reviews the ranked shortlist, filters by threshold if needed, compares top candidates side-by-side, and exports the final shortlist for client presentation.
Step 7 — Database Update	All candidate profiles and scores are stored in DynamoDB. As new JDs are posted, stored candidates are automatically re-evaluated — making the talent pool smarter over time.

Challenges & Solutions

Resume format diversity	Resumes arrive in PDF, Word, scanned images, and plain text. AWS Textract handles all formats including scanned PDFs — ensuring no resume is unreadable by the system.
Vocabulary mismatch	Candidates and JDs use different terms for the same skill. Semantic NLP embeddings via Amazon Comprehend identify conceptually similar terms, preventing valid candidates from being missed.
Explaining AI ranking to clients	Clients needed to understand why a candidate ranked where they did. The platform generates a transparent score breakdown per candidate — every ranking decision is explainable.
Scaling for peak volume	Recruitment campaigns can generate hundreds of applications overnight. AWS Lambda’s serverless model scales to process any number of resumes in parallel with no infrastructure provisioning required.
Bias in recruitment	Automated ranking risks encoding historical bias. The scoring model is built on skills and experience alignment only — personal identifiers are excluded from all scoring calculations.
Stale candidate database	Stored resumes quickly became outdated if not re-evaluated. The platform automatically re-scores all existing candidates against each new JD — keeping the talent pool continuously active.
Multi-industry keyword sets	HR Mind operates across 8 industry verticals — each with distinct terminology. The keyword taxonomy is industry-aware, with sector-specific skill libraries built for each vertical.

Key Benefits

• Resume screening time reduced by up to 90% — what took recruiters days now takes the AI minutes
• Objective, explainable candidate ranking — every shortlist position is backed by a transparent score breakdown, not recruiter intuition
• Semantic keyword matching eliminates vocabulary-based false negatives — valid candidates are no longer missed because they used different terminology
• Bulk processing on AWS Lambda — 100+ resumes processed in parallel with no performance degradation or infrastructure cost
• Continuous talent pool intelligence — stored candidate profiles are automatically re-evaluated against every new JD, turning the database into a living, searchable asset
• Faster time-to-shortlist — from days to minutes — enabling HR Mind to deliver shortlists to clients faster and win more competitive mandates
• Multi-industry ready — keyword taxonomies cover all 8 of HR Mind’s industry verticals, with recruiter-adjustable profiles per role type
• Bias-mitigated ranking — scoring based solely on skills, experience, and JD fit — personal identifiers excluded from all ranking calculations
• Strapi CMS deployed in a VM.
• Client-ready shortlist exports — formatted PDF or CSV shortlists with match scores and summaries, ready for client presentation in one click

Implementation Approach

Peritos Solutions delivered the platform in phases, with HR Mind recruiters involved at every stage to validate AI output quality against their own domain expertise:

Phase 1 — Discovery	Recruitment workflow analysis, JD structure review across 8 industry verticals, resume format audit, keyword taxonomy design, AWS architecture design
Phase 2 — Data Pipeline	AWS S3 ingestion setup, Textract PDF parsing pipeline, resume text extraction and cleaning, DynamoDB schema design for candidate and JD profiles
Phase 3 — NLP & ML Models	Resume classification model training, keyword extraction pipeline (Amazon Comprehend + custom), semantic embedding layer for vocabulary mismatch handling
Phase 4 — Matching Engine	JD keyword extraction, weighted scoring algorithm, semantic matching layer, composite ranking engine, skills gap analysis module
Phase 5 — Dashboard	Recruiter interface build — React.js frontend, API Gateway integration, bulk upload UI, ranked shortlist view, candidate comparison, export functionality
Phase 6 — Testing & Tuning	Recruiter validation of AI shortlists vs manual shortlists — model tuning to align AI output with HR Mind domain expertise; bias audit
Phase 7 — Go-Live	Production deployment on AWS, CloudWatch monitoring setup, recruiter training, hypercare support period, ongoing model improvement pipeline

Support & Next Steps

Peritos Solutions manages the AWS environment as an ongoing cloud partner — monitoring costs, model performance, and system stability. The AI models improve continuously as more resumes and JDs are processed through the platform.

Planned next phase enhancements:

• Video interview analysis — AI assessment of candidate video interviews, scoring communication skills and cultural fit signals
• Candidate outreach automation — automated, personalised outreach to top-ranked candidates via email and LinkedIn based on match scores
• Salary benchmarking integration — match scores combined with HR Mind’s salary analysis data to provide candidates with market compensation intelligence
• Real-time job board integration — automatic ingestion of applications from job boards (LinkedIn, Indeed, Naukri) directly into the AI pipeline
• Client portal — direct client access to AI-ranked shortlists with the ability to provide feedback that further trains the model to client-specific preferences
• Multi-language resume support — NLP pipeline extended to process resumes in French, Chinese, and other languages for HR Mind’s global operations

About the Client

Yorker Limited is a New Zealand-based sports-technology company founded to solve one of cricket’s most persistent problems — the inability of grassroots and amateur players to properly track and manage their bowling workloads. Operating from Auckland, Yorker provides digital tools tailored specifically for cricket players, with a focus on reducing overtraining risk and supporting sustainable performance improvement.

Key business drivers included:

• Providing bowlers with accessible, AI-driven training guidance regardless of access to professional coaching
• Enabling coaches and team managers to monitor cumulative bowling loads and prevent overuse injuries
• Delivering personalised speed improvement recommendations based on individual player data
• Building a scalable, cloud-native platform capable of growing with the user base across New Zealand and beyond

Project Background

Cricket bowlers — particularly at grassroots, club, and academy levels — lack access to the sophisticated workload management tools available to professional teams. Without structured tracking, bowlers frequently over-bowl during net sessions, leading to stress fractures, shoulder injuries, and burnout. Coaches rely on manual scorebooks and memory rather than data. Yorker was conceived to democratise performance science for cricket.

Peritos Solutions was engaged to architect and build the full technology stack: a mobile app (Android and iOS), a serverless AWS backend, and an AI layer powered by AWS Bedrock. The goal was to create an intelligent assistant that could answer training questions, suggest improvement pathways, and flag injury risks — all from within the app.

Requirements

• A mobile application for Android and iOS allowing bowlers to log net session bowling loads by delivery type and intensity
• AI-powered training suggestions using AWS Bedrock, providing personalised speed improvement drills based on player history
• Injury prevention intelligence — detecting when cumulative loads approach dangerous thresholds and alerting the player or coach
• Natural language Q&A interface powered by a foundation model, enabling users to ask questions like “How can I bowl faster?” or “Am I at risk of injury?”
• Secure user authentication via Amazon Cognito with individual player profiles and coaching team access
• Serverless, cost-efficient infrastructure using AWS Lambda and API Gateway to scale with demand
• Push notification capability via Amazon SNS for load reminders and AI-generated insights
• Full observability and alerting via Amazon CloudWatch

Solution Overview

The Yorker platform is built on a fully serverless AWS architecture, with AWS Bedrock at the heart of its AI capabilities. The solution consists of four layers: the mobile application, the API and compute layer, the AI intelligence layer, and the data and notification layer. 

 

Technology & Architecture

Mobile Platform	Android (Google Play) · iOS
AI Engine	AWS Bedrock – Foundation Model (LLM)
Compute	AWS Lambda (Serverless Functions)
API Layer	Amazon API Gateway (REST & WebSocket)
Authentication	Amazon Cognito (User Pools & Identity Pools)
Database	Amazon DynamoDB (NoSQL – Player & Session Data)
Storage	Amazon S3 (Training Data, Media Assets)
Notifications	Amazon SNS (Push Notifications to Mobile)
Observability	Amazon CloudWatch (Logs, Metrics, Alarms)
AI Prompt Design	Peritos Solutions Custom Prompt Engineering
Architecture Style	100% Serverless – No Managed Infrastructure
Region	AWS ap-southeast-2 (Sydney)

Scope & Feature List

Bowling Load Tracker	Players log each net session by delivery type (pace, swing, spin), intensity, and volume. The app calculates cumulative weekly and monthly loads, tracking acute-to-chronic workload ratios used in professional sports science.
AI Speed Improvement Suggestions	AWS Bedrock analyses the player’s logged pace data and bowler profile to generate personalised speed improvement plans — including strength drills, run-up adjustments, and release technique recommendations.
Injury Risk Intelligence	The AI monitors bowling load trends and compares them against established thresholds. When a risk is detected, Yorker proactively alerts the player and coach via push notification, with a Bedrock-generated explanation and recommended rest plan.
Natural Language Q&A	Players and coaches can type or speak questions such as ‘How do I improve my pace?’ or ‘Is my workload safe this week?’ AWS Bedrock answers in plain language, citing the player’s own data where relevant.
Coach Dashboard	Team coaches can view aggregated load data across all squad bowlers, receive AI-generated risk flags, and download session reports — giving coaching staff the oversight previously only available at elite level.
Push Notification Reminders	Amazon SNS delivers AI-triggered nudges and reminders: log your session, rest day recommended, review your weekly plan — keeping players engaged and safe between sessions.

AWS Bedrock Integration Detail

AWS Bedrock is the intelligence layer of Yorker. Lambda functions invoke Bedrock’s foundation model API, passing structured player data and a custom-engineered prompt that contextualises the AI response to cricket-specific training science. Peritos Solutions developed a prompt library covering speed improvement, load management, injury prevention, and performance analytics — each prompt template retrieves the relevant player context from DynamoDB before calling Bedrock.

The Bedrock integration handles four primary AI use cases:

• Speed Improvement: Player enters current pace figures; Bedrock returns a structured 4-week training plan with specific drills, strength exercises, and technique cues personalised to the bowler’s style.
• Injury Risk Q&A: Player asks “Am I bowling too much?” and Bedrock analyses their current acute-to-chronic ratio and provides a risk rating (green/amber/red) with an explanation and recommended action.
• Training Load Questions: Open-ended natural language questions about periodisation, recovery, rest days, and session structure are handled by Bedrock with cricket-specific context injected via prompt engineering.
• Performance Insights: At the end of each week, Bedrock generates a narrative performance summary comparing the player to their own historical averages and recommending adjustments for the following week.

Implementation Approach

Peritos Solutions followed an iterative, sprint-based delivery model across 12 weeks:

• Weeks 1–2: Requirements workshops, AWS architecture design, Bedrock model selection and prompt strategy, Cognito user pool setup, DynamoDB schema design
• Weeks 3–4: API Gateway and Lambda scaffolding, Cognito authentication flows, mobile app shell (Android & iOS), core bowling load data model
• Weeks 5–6: Bowling load tracker UI, DynamoDB integration, session logging, acute-to-chronic workload calculation engine
• Weeks 7–8: AWS Bedrock integration — speed improvement prompt, injury risk engine, natural language Q&A endpoint, Lambda prompt orchestration layer
• Weeks 9–10: Coach dashboard, SNS push notification integration, CloudWatch observability setup, performance insights Bedrock module
• Weeks 11–12: End-to-end testing, UAT with Yorker team, Google Play submission, production go-live, hypercare support

Challenges & Solutions

Challenge: AI Prompt Accuracy	Initial Bedrock responses lacked cricket-specific terminology and context. Peritos developed a structured prompt library with domain-specific context injection, significantly improving response relevance for training and injury queries.
Challenge: Load Calculation Complexity	Calculating acute-to-chronic workload ratios required handling irregular logging patterns and missing data. A robust Lambda function with rolling window calculations and data imputation logic was implemented.
Challenge: Mobile Offline Support	Players often practice in areas with limited connectivity. Local session caching was implemented in the mobile app, with DynamoDB sync triggered when connectivity is restored.
Challenge: Bedrock Latency	Initial AI response times were too slow for a smooth user experience. Lambda function optimisation, connection reuse, and streaming response patterns were implemented to reduce perceived latency.
Challenge: Personalisation at Scale	Each player requires different AI context. A DynamoDB-backed player profile system was built to inject individual history, bowling style, and goals into every Bedrock prompt — personalising responses at scale.

Benefits to the Client

• AI-powered training suggestions now accessible to grassroots bowlers — democratising elite-level performance science
• Real-time injury risk monitoring reduces the likelihood of overuse injuries through proactive, data-driven alerts
• Coaches gain team-wide visibility of bowling loads without manual tracking, saving time and improving player welfare
• 100% serverless architecture means Yorker scales automatically with the user base at minimal infrastructure cost
• AWS Bedrock eliminates third-party AI subscription costs — all AI compute is metered and cost-effective on AWS
• Secure, privacy-compliant user data management via Amazon Cognito — no player data shared with third parties
• Push notifications via SNS keep players engaged between sessions and reinforce consistent training habits
• Full observability via CloudWatch ensures the team can monitor usage, debug issues, and maintain SLA targets

Support & Next Steps

Peritos Solutions provided two weeks of hypercare post-launch, monitoring Lambda function performance, Bedrock API response quality, and DynamoDB throughput. The Yorker app is now live on the Google Play Store (com.yorker) and is actively being used by cricket players and coaches across New Zealand.

Planned next phases include:

• iOS App Store release (in progress)
• Bedrock fine-tuning with real Yorker user data to further improve suggestion quality
• Video analysis integration — allowing Bedrock to analyse bowling action video clips for technique feedback
• Team/club management features — enabling clubs to manage multiple players and compare squad load data
• Integration with wearable devices for automated session detection and delivery counting

Executive Summary

About Client

The client, Yorker, is focused on leveraging technology to address the challenge of tracking and managing cricket bowlers’ net practice bowling loads. Recognizing the risk of overtraining and injuries from improper tracking, therefore, Yorker aims to provide a digital solution tailored for cricket players. In addition, An AWS Custom Application for Yorker empowers bowlers to automate session recordings, create personalized training plans, and monitor progress effectively. The app also fosters a sense of community by enabling interaction, knowledge sharing, and participation in skill-building challenges. The project is being executed in multiple phases, beginning with a Minimum Viable Product (MVP) to establish a strong foundation for future improvements. Yorker’s commitment to innovation and user-centric design reflects its dedication to transforming how athletes manage their training and optimize performance while minimizing injury risks.

Project Background – Enhancing Cricket Training through Digital Bowling Load Management

The Yorker mobile app project addresses a major challenge for cricket bowlers: accurately tracking and managing their bowling loads during net practice. Without proper tracking, bowlers risk improper training regimens, leading to overtraining and injuries. The Yorker app offers a digital solution that automates session recordings, capturing key metrics like delivery count, types of deliveries, and intensity levels. Additionally, the app allows bowlers to create personalized training plans, track progress, and receive real-time alerts to avoid overexertion. By leveraging technology, this initiative not only helps reduce injury risks but also fosters a sense of community. Bowlers can share experiences, learn from experts, and engage in skill-enhancing challenges. Ultimately, the app aims to optimize performance while ensuring bowlers train safely and efficiently, revolutionizing the way athletes manage their training.

Scope & Requirement for AWS Custom Application For Yorker

Scope: The first phase of the Yorker mobile application focuses on developing a Minimum Viable Product (MVP) to establish a strong foundation. Specifically, this phase will deliver core functionalities to allow cricket bowlers to start tracking their training sessions and managing their profiles. The scope includes:

• User Authentication: Secure login and registration functionality for bowlers.
• Profile Management: Basic user profile setup, including personal details and preferences.
• Bowling Record Tracking: Automated entry for recording bowling sessions, including delivery count, types, and intensity.
• Basic Reporting: Simple reports summarizing bowling loads to help users monitor their progress.

Requirements:

• Mobile App Development: We will develop the front end using React Native to ensure cross-platform compatibility on iOS and Android.
• Backend Services: Built using .NET with RESTful APIs for data communication.
• Database: RDS Aurora PostgreSQL for structured data storage of user profiles and bowling records.
• CI/CD Pipeline: Set up Continuous Integration/Continuous Deployment processes for efficient development and release.
• User Interface Design: Intuitive and user-friendly UI aligned with branding, focusing on easy data entry and report viewing.

Implementation

Technology and Architecture for AWS Custom Application For Yorker

Read more on the technology and Architecture we used for AWS Custom Application Development

Technology

• WAF, API Gateway, Lambda Functions, RDS, S3, CloudWatch, Secrets Manager

Scalability

• The app is designed to run on serverless services, allowing automatic scaling based on usage.

Integrations

The application leverages RESTful APIs for smooth data transfer between the front end and back end, facilitating user authentication, session tracking, and profile management. Future integrations may include cloud-based analytics and third-party push notifications to enhance user engagement.

Cost Optimization

Peritos helped optimize costs for Yorker by designing an efficient AWS architecture using auto-scaling, right-sized instances, and serverless technologies. With tools like AWS Cost Explorer and Trusted Advisor, we continuously monitored and reduced spending. Automation through CI/CD pipelines and code optimization further enhanced performance while lowering operational costs.

Backup and Recovery

A robust backup strategy, using Amazon S3, prevents data loss, while automated recovery processes ensure quick restoration in case of failure.

Features of AWS Custom Application For Yorker

• Automated Bowling Session Tracking Capture and record each bowling session, including the number of deliveries, delivery types, and intensity levels, thus providing players with a detailed log of their training activities.
• Personalized Training Plans Create and customize training plans tailored to individual fitness levels and goals. Furthermore, Players and coaches can adjust these plans based on real-time performance data to optimize training regimens.
• Progress Monitoring & Alerts Track progress against predefined plans, with visual dashboards and alerts to notify users of deviations that may lead to overexertion or injuries.
• User Profile & Simple Reporting Maintain a personalized profile to store training history, generate basic reports on bowling performance, and gain insights to improve overall training effectiveness.

Challenges with AWS Custom Application For Yorker

• Accurate Data Capture & Tracking Ensuring the app reliably records detailed bowling metrics like delivery type, count, and intensity without manual errors poses a challenge, especially in a real-time sports environment.
• Scalability & Performance As user adoption grows, maintaining app performance and scalability will be critical, particularly during peak usage times. Designing a backend that can handle large volumes of data efficiently is essential.
• User Engagement & Retention Encouraging consistent use of the app among bowlers can be challenging. Building features that foster community interaction, personalized plans, and gamified challenges will be crucial to retaining users.
• Cross-Platform Compatibility Delivering a seamless user experience across both iOS and Android devices requires rigorous testing to address device-specific issues, screen resolutions, and performance variations.

Support

As part of the project implementation we provide 2 months of Ongoing extended support. Additionally, this also includes 20 hrs a month of development for minor bug fixes and a SLA to cover any system outages or high priority issues.

Next Phase

We are now looking at the next phase of the project which involves:

• Ongoing Support and adding new features every Quarter with minor bug fixes
• Social & Community Building Features