I build AI systems that survive contact with production.
Agent evaluation · Model behavior · Production machine learning
- Arka: Config-driven synthetic data for LLM training, with Evol-Instruct, IFD scoring, deduplication, and PDF ingestion.
- Humanize RL: RL environments and reward modeling for more natural AI-generated text.
- Medha: Local-first SQL IDE for flat files with AI, DuckDB, FastAPI, Vite, and LangGraph.
- Entity Resolution POC: Embedding-based entity resolution at 500M-record scale.
- Pi Agent Extensions: Extensions for sessions, user prompts, handoffs, and coding-agent workflows.
- Sangam: Self-hosted document server where humans and identified AI agents work with ordinary files through one API.
- Text Watermarking Lab: Experiments with watermarking and provenance for generated text.
- Can Dense Retrieval Beat BM25 for Entity Resolution?
- File-Based Memory Is a Terrible Idea That Somehow Works
- Stop Renting Your Workflow: Building a Custom AI Coding Agent with Pi
When the p-value is finally below 0.05.





