At Inscripta AI, I built production LLM evaluation and RAG systems. Hybrid search and custom document parsers raised retrieval accuracy above 80% on real queries, and I developed frontend features for a multi-agent planning system.
At Drevan, I built Control Room, an agent runtime with expiring permissions, spend checks, crash-resumable runs, and human approval gates. I also developed Maya, a public real-time voice agent in 16 languages, with an ONNX turn detector reaching 162 ms p90 on 2 vCPUs.
I’ve also built a rooftop-solar marketplace and open-source RAG work with cited answers and abstentions. My projects include evaluating LLMs and agents against holdout data, latency, and cost.
