At Synctom, I architected Haven, a real-time voice AI system streaming 16 kHz audio over duplex FastAPI WebSockets to Next.js. I engineered two voice engines and reduced production first-audio latency from 3.0–3.7 seconds to 1.6–2.1 seconds.
On the Haven project, I launched LLM and TTS from Flux’s eager end-of-turn signal and ran RAG over partial transcripts, cutting planning time from 616 ms to 1 ms. I also implemented an interruption judge and an LLM-free crisis circuit, validated with 78 pytest tests and 17 red-team cases.
I built Hallucinot to score claims against source passages, reaching 92.8% accuracy and routing misses to human review. At CloudSec Network, I built and secured AWS infrastructure and automated multi-tier deployments with Terraform, reducing provisioning time by 80%.
