Projects
Built, shipped, and thought through.
Work from my jobs, things I built on my own, and systems I have designed on paper. Each is labelled, so you always know which is which.
Work
Built at a job.
Claims API test framework
The Java 17 + TestNG + Rest-Assured framework that tests health-insurance claim, enrolment and partner APIs.
Log-to-test microservice
Took real HTTPS traffic from ELK logs, turned it into curl commands, and ran them as API tests.
Real-time backend with a chatbot
Node.js and Express services with MongoDB, JWT login, live messaging over Socket.io and a Dialogflow chatbot.
Built
Built on my own.
QueryLite — a SQL database from scratch
A relational database engine: SQL parser, B+ tree indexes, a cost-aware query planner, three join algorithms and transactions. Runs live on 22,000 rows.
Raft consensus, implemented and visualised
The consensus algorithm behind etcd and Kubernetes: leader election, log replication and commits — tested under random crashes and network partitions.
Live collaborative editor (CRDT)
Google-Docs-style real-time editing with no server: a CRDT keeps every copy in agreement, even after offline edits. Syncs live between browser tabs.
Connect to Nature
A farm-stay marketplace for the Kokan and Nashik regions — one app, two portals, in English, Hindi and Marathi.
Rakta-Setu
Sends blood reports to patients on WhatsApp in Marathi, with a voice web page that explains every result.
Generic Dynamic Journey Builder
Turns a Chrome DevTools HAR export into a repeatable API test — it pulls out values, chains the requests and checks the results.
JSON-Driven API Sequencing Engine
Chains API calls from a JSON config, with retry, polling, live SSE output and a visual builder.
Wildlife Conservation Analysis
Counts animals in images and video and classifies species against the IUCN Red List. Copyright-registered.
Ask-me assistant
Ask anything about me. It searches this site first, then Gemini writes a short answer from those pages only — always with the source.
Guardrail game
Catch prompt-injection attacks and let safe prompts through — an arcade round, then a slower round scored on precision and recall.
LangGraph: wire the agent
A puzzle game that teaches how LangGraph agents work: connect the steps, run the tests, fix the loop.
Langfuse: trace detective
Six user complaints, six traces. Find the step that caused each one — and learn how to read an AI app’s traces.
System design simulators
Rate limiter, cache stampede, circuit breaker and capacity planner — watch each idea work instead of reading about it.
This website
A fast, multi-page site with a spring-physics background, light and dark themes, and every tool and game running in the browser.
Designs
Worked out on paper. Not shipped.
Agentic claims triage
A plan → tools → check loop that reads a claim, picks which checks apply, and asks a human when evidence is thin.
Document extraction flow
Scanned bills to structured JSON, with schema checks and a review queue for anything uncertain.
LLM gateway
One endpoint in front of several models: routing, fallback, caching and a spend limit per team.
Prompt release pipeline
Prompts treated like code: versioned, reviewed, rolled out slowly, and rolled back in one step.
Support ticket auto-reply
Sorts a ticket, finds the right runbook, drafts a reply, and hands it to a person when unsure.
Prompt regression tests
Golden test cases that run in CI, so a prompt or model change cannot quietly break a right answer.
RAG retrieval tests
Tests the “find the right document” half of RAG on its own, because most bad answers start there.
Hallucination checker
Splits an answer into single claims and checks each one against the source.
LLM load and cost test
Load-tests an LLM feature like any other service: response times, tokens per user flow, and the breaking point.
API contract drift check
Compares OpenAPI specs between environments and blocks the deploy on a breaking change.
Grounded knowledge assistant
An assistant over runbooks and API specs that keeps every answer attached to its source.