Applied AIResearcher & Engineer

Built a lightweight on-device ASR system via knowledge distillation from a Fast Conformer RNN-T teacher, achieving a 32.7% relative WER improvement — the highest reported for Conformer RNN-T + KD.

Proposed MoPE-LoRA, a PEFT framework that routes Conformer encoder frames to six manner-of-articulation LoRA experts via hybrid phoneme-supervision and learned acoustic gating — no accent labels required. On L2-ARCTIC, it achieves 10.43% WER, outperforming Full FT (12.80%) and Single LoRA (11.33%), with 12.3% relative zero-shot improvement over Single LoRA across unseen accents.

Implemented a full in-browser ASR pipeline in JavaScript using ONNX Runtime Web, covering audio capture, log-mel spectrogram extraction, encoder-decoder inference, and text decoding — entirely client-side with no server dependency.

Technical lead on a Bill & Melinda Gates Foundation-funded project. Deployed a RAG-based SRMH chatbot serving 5,000+ Nepali users and fine-tuned LLMs for low-resource Nepali generation.

Technical lead on a UNDP-funded project. Built a multilingual RAG chatbot to digitize citizen charters, deployed across two municipalities after surveying 3,000+ citizens.

Built SLAM-based localization APIs enabling persistent AR on mobile devices, implemented RGBD and Monocular Visual SLAM systems, and developed the ROS package Kachuwa during a research internship at NAAMII.