Three moments, one thread
From hardware efficiency and low-level optimization to running production platforms and shipping enterprise systems.
Journey & Timeline
Curiosity at the hardware level
100 Days of CUDA — GPU programming fundamentals, parallel algorithms, CNN implementation, and Triton kernels, built to understand how models actually run underneath.
Three papers in under a year
On-device LLM optimization (Fragile Mastery, arXiv:2503.22698), interpretability via inverse reasoning (Thinking About Thinking, arXiv:2507.00092), and medical visual question answering (LoRA-Enhanced PaliGemma, MediaEval 2025).
Shipped, in production
Enterprise KYC vision systems (Helios), multi-agent hosting infrastructure (Foundry), and localized speech technologies for underserved languages (Nepali & Maithili speech AI).
Flagship Deployments
Helios
LIVE ENTERPRISE PLATFORMAI-powered KYC platform with passive liveness detection and document extraction, co-built and deployed by enterprise fintech clients including Khalti.
Foundry
ACTIVE INFRASTRUCTUREEnterprise-grade AI agent hosting and execution infrastructure, built to run complex multi-agent workflows at scale with low latency.
Nepali & Maithili speech AI
IN PRODUCTIONMultilingual TTS, ASR, and Indic↔English machine translation, built for languages with little existing AI tooling.
Additional Codebase Records
Fine-tuned Chatbot
Streamlit app using BART-base fine-tuned on Bitext customer support dataset, backed by SQLite3.
Web Agent
Web agent built from scratch using Hugging Face smolAgents, hosted on Hugging Face Spaces.
Brain Tumor Detection
CNN-based detection from medical images with PyTorch and responsive UI representation.
Ecolog
Hackathon project, collaborative tech to prevent forest fires and poaching in conservation areas.
NeuroNav
EEG-based wheelchair control prototype, capturing brain waves to assist mobility.
arXiv Summarizer
Flask app summarizing research papers with LaTeX equation rendering support.
Interactive Web Agent
Fellowship.AI major project focusing on GUI-based web navigation and reinforcement models.