Work · Narasimha Mujumdar

What have I built?

data platform · verification

problem: Record counts can match while the contents differ. There was no way to show that a migration had moved everything.
built: A tool that derives a data graph for each service and checks the migration against it using content hashes. An AI model interprets the infrastructure code; a deterministic engine derives the schema from the data.
value: Completeness could be shown service by service, and differences were found during the migration rather than later.

data migration · graph modelling · content hashing · AI-assisted

identity · SCIM 2.0

problem: Telephony users were managed in Entra ID. They also needed accounts in Amazon Connect, kept in step as people joined, moved and left.
built: A custom SCIM 2.0 API, written from scratch, connecting Entra ID to AWS IAM, with SAML 2.0 federation for sign-on.
value: The cutover and the ongoing user lifecycle were both automated. I designed and built it.

SCIM 2.0 · SAML 2.0 · Entra ID · Amazon Connect

digital health · resilience

problem: A nurse-on-call service records every call. Recordings must not be lost, and the existing pipeline was failing too often.
built: A TypeScript and AWS pipeline using SQS and Lambdas, where work is queued and retried instead of dropped. Later re-architected to improve fault tolerance.
value: Failure rates dropped. Under heavy load the system slows down rather than losing recordings.

TypeScript · AWS Lambda · Amazon SQS · fault tolerance

digital health · self-initiated

problem: A contact flow can break while the infrastructure metrics still look normal. These failures were usually reported by callers.
built: A framework that places automated calls, transcribes them with Whisper running locally, and compares the result against what the flow should say.
value: Failures are detected automatically, which shortened the support team's reaction time. The models run locally, so audio stays in-house.

Whisper · Hugging Face · Twilio · Amazon Connect

automotive · Go

problem: Dealers at a top-selling car company wanted to offer rentals lasting months. Availability, pricing and allocation all work differently over that length of time.
built: Go APIs on AWS handling the booking and availability logic across dealers, with payments centralised through Stripe. I worked on it from requirements through to deployment.
value: A long-term rental product the dealer network could offer to customers.

Go · AWS · Stripe · REST APIs

retail · performance

problem: The core shopping APIs of a national Australian retailer were slow under load, and the causes were spread across legacy code paths.
built: Profiled the application with Dynatrace to find where the time was being spent, then refactored those paths.
value: API response times improved by 20% to 60%.

performance profiling · Dynatrace · refactoring

agentic AI · RAG

problem: Retrieval systems report their own sources, and those citations are sometimes wrong.
built: Semantic search over a large OCR'd corpus with agentic retrieval, and a separate deterministic engine that matches every citation back to the source text.
value: Citations are checked against the source rather than taken from the model's own report.

Python · RAG · agentic tool-use · embeddings

telephony failover. Amazon Connect set up as failover for a Twilio contact centre, removing a single point of failure.
Tesco, as solution architect. Added circuit breakers to the grocery storefront to improve uptime, and re-architected Transport & Tracking on SOA principles for the global divisions.
monolith to microservices. Order management and voucher distribution moved into decoupled Node.js services.
real-time score delivery. Node.js and Redis pushing live match scores to a UK wagering platform.