|
Lead AI Engineer. Chennai. Hand me a vague business problem and I hand back a deployed system. Sitting with the business team to figure out what they actually need, drawing the architecture, writing most of it, shipping it, then keeping it alive in production. That whole arc is the job and I like all of it. Voice agents are where I have gone deepest: real-time speech, turn-taking, barge-in, telephony, orchestration. Around that sits everything else an AI product needs to actually exist. Agent graphs, retrieval, fine-tuning, the API, the interface, the infrastructure. None of those layers stay in their lane. That is the interesting part. |
|
latency is a feature. users feel milliseconds before they notice quality.
the interesting bugs live at the boundaries between services, not inside them.
an eval you do not run before every release is a decoration.
a demo that works once is not a system.
if it is not deployed, it is not done.
Most of what I build is private. What is not:
spectrahush spectral subtraction that runs inside the audio callback. no weights, no gpu.
denoisestream websocket mic denoising. swap the backend without touching the transport.
aws-ai-agents agentic cloud cost analytics on bedrock agentcore.
ai-engineer-agent multi agent system that reads a repo and tells you what is wrong with it.
python · typescript · fastapi · next.js · postgres · redis · docker · kubernetes
langgraph · pipecat · livekit · webrtc · faiss · unsloth · vllm · langfuse