Machine learning engineer working on perception pipelines for robotics: segmentation, 3D mesh, pose and PII redaction across egocentric, stereo and fisheye video. Before that I built synthetic data pipelines for enterprise agents, shipped a voice assistant to npm and spent three years finding critical vulnerabilities in software people use every day.
What I am building now
Robot CI for learned robot policies. It helps robotics labs check whether a policy change makes a real difference before testing it on a robot or showing it in a demo. Its six modules check that the deployed policy matches the evaluated one, report rates with uncertainty intervals, track episode records and fleet metrics, detect drift in the cell and arm, judge outcomes and guard execution at runtime. The contract checker blocks execution when checks fail. The outcome judge abstains when it is uncertain.
I publish measurements on public data with the scripts that produced them. That includes runs that contradicted my earlier claims. One finding corrected my own tool: all ten failure detectors performed at chance on actual robot task failures despite catching injected execution faults almost perfectly. Nine of those detectors came from recent literature.
DESCRIBE SELECT * FROM (...) statement they never wrote. Errors during execution now omit that location block and keep the diagnostic text. Parse errors retain it because they refer to the caller's own SQL.Ready to train logs and usable models for physical AI in dark stores, factories and warehouses.
A stablecoin payments app built for B2B businesses and remittance with a waitlist of businesses and individuals.
Major subjects: Computer Science, Maths and Physics.
Percentage: 85.6% out of 100.
Physics, Chemistry, Maths and Computer Science.
Percentage: 80% out of 100.
Physics, Chemistry, Maths and Computer Science.
JavaScript, Python, React, HTML, CSS, SQL, Linux, APIs, Node.js, Git and MongoDB.
Hindi, English, Awadhi and Bhojpuri.
Chess, football and the internet.
Writing something.