At CONVIN.AI, I build multilingual ASR for BPO call-center use cases, covering eight Indian languages and English. I developed Transformer-Transducer models with a dual-head decoder for native-script and phoneme prediction, and fine-tuned a Hinglish model to state-of-the-art performance.
I extended Qwen3-ASR-1.7B’s vocabulary for Indic languages and implemented two-stage fine-tuning on 90k hours of speech data, achieving 15.31% corpus WER on the VISTAAR benchmark. I also served a Gemma 3 + LoRA soft-prompt ASR decoder on vLLM using a custom virtual-ID audio-injection scheme.
At Samsung Research Institute-Delhi, I developed a speech-denoising system integrated with visual displays to improve the TV viewing experience for ASD patients. At IIT Kanpur, I integrated a persistent-memory tier into the WRF weather model, cutting output-write time from 8.5 seconds to 4.0 seconds versus HDD.
