I build and debug real-time voice AI for a state electricity utility’s 1912 complaint helpline (TGSPDCL). The agent speaks Telugu, Hindi and English over the phone, collects the caller’s consumer number, looks up accounts and feeder outage status through backend APIs, and books complaint tickets.
My work spans the full pipeline: voice activity detection and turn-taking, language identification and routing for code-mixed Telugu/Hindi/English speech, LLM-driven dialogue with strict tool calling (Pipecat flows, vLLM serving Gemma 4 31B on an NVIDIA H200), Indic text-to-speech (Piper, VibeVoice, IndicF5 voice cloning), speaker verification (ECAPA-TDNN), and the Redis/PostgreSQL/monitoring services around it.
I spend as much time on measurement as on code: I audited 1,000+ real calls to find where callers actually fail and redesigned the consumer-number collection step from that data, chose the production LLM by measuring Telugu tokenisation cost and replaying real turns, traced a production outage to its root cause, and quantified GPU thermal throttling’s effect on response latency. I report honestly what is proven and what is not.
Before this I spent five years in client-facing data and analytics roles (SQL, Python, Azure Data Factory, Power BI).
Looking for roles in voice AI, applied LLM engineering, or forward-deployed AI engineering, remote or Hyderabad.
