•Engineered an agent harness to replace a third-party framework, with sub-agents, parallel tool calls, sandboxed code execution, versioned file workspaces, and async long-term memory of user preferences.
•Raised prompt-cache hit rate from 30% to ~100% by keeping prompt prefixes byte-stable across turns; cut tools loaded per request by 60% and per-request tokens by 40% with progressive disclosure of tools and skills.
•Rebuilt the drafting system from fixed DAG pipelines into tool-calling agents with programmatic validation and reusable skills, improving eval pass rate by 60% and cutting end-to-end latency by ~37%.
•Set up a self-hosted evals and observability stack on Langfuse and ClickHouse to trace agent runs and track eval results.
•Shipped end-to-end product features across the Microsoft Word add-in, Chrome extension, mobile app, and web frontend.
•Shipped a production RAG system with BM25, dense retrieval, query expansion, reranking, and conversational memory, improving retrieval relevance by 40% and user engagement by 25%.
•Created an LLM prompt-optimization pipeline with chain-of-thought and few-shot examples, improving regulation and urban-planning document analysis accuracy by 35%.
•Cut response latency by 78% and memory footprint by 71% via connection pooling, leaner data structures & pipeline refactoring.
•Automated evaluation with an LLM-as-a-Judge framework using GPT-4.1, reducing manual annotation effort by 80%.
•Built FastAPI services to automate scraping of 10,000+ Indian court case files, improving legal data accessibility by 60%.
•Designed a LangGraph multi-agent system for contract analysis, identifying adversarial clauses with 85% precision.
•Delivered a production RAG pipeline with hybrid search for a legal Q&A chatbot, improving answer relevance by 40%.
•Implemented document preprocessing and chunking workflows for long-form legal text, improving retrieval consistency and reducing noisy context in downstream contract and case-law analysis.