At REVRAG.AI, I built and operate a production real-time voice pipeline integrating ASR, LLM, TTS, and VAD on LiveKit. I reduced end-to-end conversational latency by 40% and built interruption-handling controls that reduced the interruption rate by 20%.
I also fine-tuned Gemma 27B with LoRA for phone-based BFSI customer onboarding and prototyped a turn-detection model using acoustic and semantic signals. I benchmarked self-hosted open-source models against proprietary APIs to document quality, latency, infrastructure, and cost tradeoffs.
At LOHUM CLEANTECH, I designed a CNN for lithium-ion cell assessment that reduced testing time by more than 95%; the method is the subject of two granted patents. I also trained battery degradation models using vehicle telemetry and deployed models on AWS, reducing inference cost from INR 0.21 to INR 0.013 per cell.
