At Kurly Brains, I lead data curation and LoRA/QLoRA fine-tuning of open-source LLMs and multimodal models for client applications, with production handoff. I also explain model architectures, evaluation metrics, and practical trade-offs to clients and team members.
At Information & Language Processing Research Lab (ILPRL), KU, I worked on Nepali–Tamang machine translation and benchmarked multimodal LLMs for code-switched Nepali–English speech recognition. My first-author paper at CHiPSAL 2026 @ LREC explored reward-guided Whisper fine-tuning for low-resource Nepali speech recognition, reducing WER from 5.60% to 4.89%.
I’ve also worked on a Nepali–English code-mixed speech corpus and Whisper adaptation, and I’m building a Newari ASR system with synthetic TTS augmentation. My projects include a drone-imagery land-cover segmentation pipeline and an open-source browser extension for Nepali transliteration, OCR, and form autofill.
