At VideoSDK.live, I lead a six-engineer team working on voice AI, transcription, real-time communication, and robotics teleoperation. I take production customer issues from rapid proof of concept through rollout.
I integrated in-house trained ASR and TTS models into the production voice pipeline and optimized inference on NVIDIA L4 GPUs, cutting end-to-end latency from 1.5 seconds to 800 milliseconds. I also architected a no-code voice agent builder, including its inference router and SIP telephony integration.
I fine-tuned Gemma 4 and built low-latency serving for language and turn-detection models using vLLM and SGLang. I contributed to the architecture and training of the Echo Omni and Echo Large models, and scaled WebRTC, SIP, and cloud recording infrastructure while reducing cloud costs.
I also shipped robot teleoperation over WebRTC with measured end-to-end latency of 50–80 milliseconds, and migrated observability to a self-hosted stack handling 400K logs per minute at half the cost. Earlier, as a Software Development Engineer Intern at IBM India Software Labs, I automated security repository reporting and built a log and metrics microservice.
