Senior Machine Learning Engineer · Tech Lead, Agentic Infrastructure
I'm a Senior Machine Learning Engineer at Fetch and work as Tech Lead for Agentic Infrastructure. I lead the infrastructure work behind production agents, from evaluation and deployment to observability and day-to-day operation. I shipped an AI Shopping Assistant to a 200K-user experiment and worked on a multi-agent matching framework that unlocked $165M in GMV with LLM-as-a-Judge guardrails at 97% accuracy.
Over the past several years I've worked across computer vision, multimodal and contrastive learning, and NLP before specializing in agentic systems. These days I'm drawn to the hard, unglamorous parts: evaluation loops that gate deploys, context that stays within a token budget, and coordination protocols for agents that outlive a session - work that shows up in the open as Agent Relay and Lazyflow.
My path started in hardware, building autonomous robots for the ABU Robocon international competition, then representation-learning research at NASA-IMPACT during my MS in Computer Science. Most of what I learn still ends up as an open-source tool.
Systems I work with
Language and multimodal models used as capabilities within larger systems.
Predetermined LLM pipelines, retrieval patterns, and structured model steps.
Stateful agents that choose tools and adapt their process to a task.
Specialized agents coordinated through roles, state, handoffs, and review.
Quality gates, traces, feedback signals, and continuous improvement.
Services, persistence, deployment, and operational systems behind AI products.
Fetch · Remote
Leading the infrastructure work behind Fetch's production agents, from evaluation and deployment to observability and day-to-day operation.
Built and shipped the AI Shopping Assistant to a 200K-user experiment, plus an end-to-end LLM evaluation system with CI/CD deploy gating that blocks releases on quality regression.
Led the Multi-Agent Matching Framework, improving match rates by 3.1-3.3 points and unlocking $165M in GMV with LLM-as-a-Judge guardrails at 97% accuracy.
NASA-IMPACT
Research in Machine Learning and Natural Language Processing at NASA's Interagency Implementation and Advanced Concepts Team.
Fusemachines Nepal - Kathmandu, Nepal
Explored and quantified product cannibalization effects for retail fashion, delivering actionable insights for merchandising strategies.
Led efforts to build an ensemble model classifying fashion products into Kate Spade's four customer segments using computer vision and SciPy optimization, achieving 62% accuracy.
Developed an in-house recommendation system using a Factorization Machine model, AWS EC2, and Lambda, achieving ranking metrics comparable to AWS's recommendation engine.
Taught data science, computer vision, machine learning, deep learning, and MLOps through Fusemachines AI Fellowship programs.
Betterhalf.ai - Bangalore, India
Built NLP systems for automated user profile generation and content moderation at a consumer matchmaking platform.
Robotics Club, Pulchowk Campus - Tribhuvan University
Designed and built autonomous robots for ABU Robocon, mentored junior members, and received a 2019 ROHM Award for robot design.
Academic journey from electronics engineering to machine learning
Huntsville, AL
GPA: 4.0 | Thesis on Spectral Deconvolution using Machine Learning
Online
Grade: B | MIT Professional Certificate Program
Lalitpur, Nepal
Senior Project: Sound Source Localization for Disaster Victim Search