Naman Gupta | India | Remote only | Contract (B2B, USD/EUR) or full-time
AI engineer focused on retrieval quality and agent reliability. I measure retrieval before changing it.
- Built and evaluated a production document-AI retrieval pipeline and LangGraph workflows.
- rag-eval-harness (open source): 20 CUAD contracts, 50 queries, 114 annotated spans, 7 retrieval configs. Clause-aware chunking kept 99.1% of answer spans intact vs 43.0% for fixed-width. Hybrid BM25 + dense with RRF scored best (MRR 0.70). The reranker did not help. 87 unit tests.
- Merged PR in Lamatic/AgentKit: EvidenceFit, compares chunking strategies against labelled evidence spans and returns a SHIP/TUNE/BLOCK verdict.
- MCPHub: MCP server testing and inspection across SSE, Streamable HTTP and stdio, with latency percentiles.
For contracts: a 5-day fixed-scope retrieval baseline on your pipeline (labelled eval set, failure patterns, prioritized fixes, harness handed over).
AI engineer focused on retrieval quality and agent reliability. I measure retrieval before changing it.
- Built and evaluated a production document-AI retrieval pipeline and LangGraph workflows.
- rag-eval-harness (open source): 20 CUAD contracts, 50 queries, 114 annotated spans, 7 retrieval configs. Clause-aware chunking kept 99.1% of answer spans intact vs 43.0% for fixed-width. Hybrid BM25 + dense with RRF scored best (MRR 0.70). The reranker did not help. 87 unit tests.
- Merged PR in Lamatic/AgentKit: EvidenceFit, compares chunking strategies against labelled evidence spans and returns a SHIP/TUNE/BLOCK verdict.
- MCPHub: MCP server testing and inspection across SSE, Streamable HTTP and stdio, with latency percentiles.
For contracts: a 5-day fixed-scope retrieval baseline on your pipeline (labelled eval set, failure patterns, prioritized fixes, harness handed over).
Stack: Python, TypeScript, LangGraph, FastAPI, Next.js, Qdrant, pgvector, Postgres, Docker.
Proof: https://namangupta.dev/production-ai
GitHub: https://github.com/namanxdev/rag-eval-harness
Email: [email protected]