Work

Watchdog agent for agent fleets. Catches policy breaks in traces, quarantines the run, patches the guardrail and confirms the fix.

Voice assistant in the Mac notch. Push-to-talk realtime speech, background sub-agents and MCP tools.

RL environment where a supervisor LLM manages five worker agents, some deceptive, some colluding. GRPO on Qwen2.5-3B.

Cited answers over 512K company docs, served to agents over MCP. +20 over BM25 on a held-out test split.

Multimodal RAG over hardware datasheets. Routes the question, retrieves figure and table crops, cites the page.

ML paper to implementation brief, with a cited Q&A agent and a skeptic pass. LangGraph, FastAPI.

RL environment where an agent resolves merge conflicts block by block. OpenEnv.

Discharge summary in, picture and voice medicine schedule out. Never changes a dose. Lambda, Bedrock, DynamoDB.

Credit appraisal engine with Celery pipelines and XGBoost risk scoring. IIT Hyderabad hackathon finalist.

Stateless Node.js WebSockets, 1,000+ concurrent connections.