Work

Built milanSathi, an in-house AI-powered astrology engine (kundli + kundli-milan) that replaced a paid third-party astrology API.

• 100% parity with the vendor on everything tested

Built an in-house feature-configuration platform (Spring Boot, Redis, MongoDB) that replaced Firebase Remote Config after it moved to a paid tier.

• 10M+ requests/day

Built AI first-message suggestions. When someone sends an interest, an async Kafka worker reads both profiles and asks Qwen (on Vertex AI) for personalised openers, so the sender can pick one instead of sending a blank interest.

Implemented Redis semantic caching to avoid duplicate LLM calls.

• 40% lower message latency

Delivered a real-time GenAI matchmaking engine using Spring Boot, Openfire (XMPP) and Aerospike, orchestrating GPT, Gemini and Qwen through a LangChain/LlamaIndex RAG pipeline on RedisVL vector search.

Led LLM integration across a Kubernetes-deployed React/Redux and Spring Boot stack, using Structured Outputs (JSON Schema) to regenerate listing data.

Designed a fault-tolerant, event-driven matching algorithm on Kafka and NoSQL stores.

Built containerized pipelines sending conversion events to Meta and Google APIs, with MongoDB aggregations and cron batches automating daily marketing feeds.

Built a tool that tailors a LaTeX resume to a job description by reordering skills only. It can never add a skill that isn't true, and it reports gaps honestly.

• 89% on an AI Engineer JD

Built an installable MCP server that reviews code diffs against a repository's own conventions, using RAG (AST chunking, on-device FastEmbed embeddings, Chroma) and a LangGraph review agent.

• 0 false positives across a 6-case eval suite

Contributed to the team's internal AI engineering assistant. I built its live-production Debug/RCA tool (reads prod logs and cross-service traces, leads with evidence, never a verdict), one-command setup, conversational follow-ups and a plan-approval gate in its agent engine.

Prototyped a telemetry-driven service that re-ranks UI components per session from clickstream signals using an LLM-scored layout policy.

Built a marketing landing page.

Revamped OTP authentication with Spring Security and JWT, tuned MySQL (InnoDB) queries and added Prometheus anomaly detection.

Standardised cross-service calls through a containerized Spring Boot API aggregator with unified headers.

Moved the profile approval workflow to an event-driven design on RabbitMQ with a central listener factory.

Trained a deep learning model (TensorFlow, PyTorch) to predict DNA methylation sites for cancer diagnosis.

Built an academic peer-doubt platform on Django REST and SQLite, with indexing to speed up study-group lookups.