AI Engineer, Bengaluru
I build production AI systems and the instruments that measure them. Real-time voice pipelines, evaluation frameworks, data infrastructure. Everything below is running or public.
user stops speaking
This is my own voice agent, measured. The overshoot is the interesting part.
auviaengine.com · github.com/priyalrathore003/auvia-engine
Production audio platform, built solo. Real-time voice agent, LangGraph orchestration, async DSP, deployed on GCP.
Findingaudio payloads were entering LLM context, producing 1.2M-token requests. Passing references instead of payloads cut inference cost ~40%.
github.com/priyalrathore003/voice-latency-harness
Open-source evaluation framework for voice pipelines. Reconstructs per-turn timelines from webhook events, reports p50/p95 per stage against a budget.
Findingpointed at a commercial platform's real webhook contract, all seven pipeline stages come back unmeasurable — each naming the specific event that's missing.
github.com/priyalrathore003/clinical-scribe
Multilingual clinical documentation for Hindi-English consultations. Every field traceable to a timestamp and verbatim quote.
Findingtwo ASR models disagreed on a drug dosage — 500mg versus 400mg. A single-model pipeline resolves that silently. This one surfaces it.
github.com/priyalrathore003/earnings-call-analysis
Measures how management speech shifts between scripted remarks and live Q&A — hedging density, response latency, speaking-rate variance.
Findingbuilt with a statistical validation layer, because the honest expected result is no signal, and reporting that rigorously is the whole point.