Senior AI Engineer at Tiger Analytics (Bengaluru). I build production AI that enterprises can trust: agents that plan and act, retrieval grounded in real data, and the guardrails, evaluation and tracing that make them safe to ship.
- LLM guardrails. Small fine-tuned models that catch PII leaks and prompt injection at low latency, with traceability from training data to the deployed model.
- Agentic AI for financial services. Multi-agent workflows, Text-to-SQL and RAG analytics, and governance (identity, human approval, audit logs) on AWS Bedrock.
- Evaluation and observability. Testing AI behaviour in CI, tracing agents with OpenTelemetry, and scoring output quality.
| Project | What I fixed or added |
|---|---|
| deepset-ai/haystack and haystack-core-integrations | Silent data corruption from shared dataclass mutation across pipeline branches; Anthropic server tools; cache-token usage for LiteLLM, Gemini and Bedrock; async document stores |
| NVIDIA/garak | Retry transient NIM HTTP errors and report the real status code |
| openclaw/openclaw | Plugin registry re-activating on every MCP connection; Linux daemon crash on bus permission errors |
| Arize-ai/openinference | MCP uninstrumentation; cost propagation to span attributes |
| openlit/openlit | Token counts and tool-call spans across providers; offline evaluations after restart |
| truera/trulens | SQL-level leaderboard aggregation; correct cost for unknown models |
| comet-ml/opik · langsmith-sdk | Streaming usage helpers; Pydantic warnings in the Anthropic wrapper |
| Azure Cosmos DB MCP samples | The first Python MCP server for Azure Cosmos DB |
| Project | What it does |
|---|---|
| BehaviorCI | CI for LLM behaviour: describe how your AI feature should behave, and it generates adversarial tests and blocks the PR when behaviour regresses. npx behaviourci demo |
| greenwash | Catches coding agents that weaken your tests and then report success. Deterministic: reads git diff, no model, no API key. |
| Native Place | Find the train to your hometown: any town in India, official 2026 timetable, fares and festival specials. Live ↗ |
| SLM Guardrails Chat | Local small language models acting as safety, topic, format and PII guards in front of a Groq-powered chatbot |
| Prompt Injection Gym | A small, reproducible playground to attack and defend a tool-using agent |
| Book Writing AI Agent | Multi-agent system that plans, writes and edits a full book |
Python · TypeScript · FastAPI · LangGraph · Strands · MCP · AWS (Bedrock, AgentCore, S3, Athena) · Azure · PyTorch · Hugging Face · MLflow · OpenTelemetry · PostgreSQL · Redis · Docker
Open to conversations about AI engineering roles and open-source collaboration.
