I trace every claim to evidence
Public pages point to approved proof, source cards, or clearly labeled representative artifacts. I don't make claims I can't back.
About
8 years building AI systems that survive production.
I've worked across computer vision, ML infrastructure, search systems, and agentic AI platforms. My focus is the part most teams skip: making AI workflows observable, evaluable, recoverable, and affordable.
I'm currently looking for Staff/Principal AI engineering roles where reliability and cost matter as much as capability.
The arc
Where I've been.
2013-2018IIT-BHU Varanasi, Dual Degree in Computer Science (9.28/10)
2016Microsoft Research intern: dialog systems and chatbots
2017UC Berkeley research intern: neural programmer-interpreters (Prof. Dawn Song). SN Bose Scholar.
2018-2019Whodat: built C++ ORB detector 20% faster than ORB-SLAM for AR products
2019-2023Osmo: CV technical lead across India and US teams. 93% → 98% worksheet recognition accuracy.
2023-2024Epic! for Kids: owned ML platform post-layoffs. 10x infrastructure cost reduction.
2025-2026Knit: principal architect for agentic market research platform. 48-72h → <1h report turnaround.
2025Kaggle top 6% globally. Open-source ML projects.
NowOpen to Staff/Principal AI systems roles.
How I work
Beliefs shaped by production.
I prefer explicit workflows
DAGs, recovery states, evals, and logs over unstructured prompt chains. If I can't debug it, I won't ship it.
I design for the next engineer
The work only lasts when someone else can understand the contract, failure mode, and evidence trail without asking me.