At WeyaAI, I built agentic voice AI serving 500K+ minutes and 3M+ calls per month at under 700ms latency. I also improved the audio noise-mixing pipeline, reducing per-frame CPU processing from 113.5ms to 0.07µs.
I reduced first-message TAT to 200ms and post-call analysis LLM cost by 70%, while increasing prompt cache-hit ratio from 40% to 90%. I also set up observability, alerting and LLM evaluations, and built a unified memory layer for voice, WhatsApp and email agents.
Previously, I built backend and AI systems for natural-language Q&A over files and SQL, agentic workflows, and reporting tools. I’ve contributed to LlamaIndex and QB64, and co-created a QB64 port of p5.js.
