Harsh - AI person, mostly
I write Python for a living, train neural nets for fun, and occasionally ship things that actually work. MS in AI at Boston; graduated April 2026 and very much open to what's next.
My Toolkit
AI & Deep Learning
Frontend & Mobile
Infrastructure
Projects
Paragrad Autograd Engine!
2026
Rocket Landing using Reinforcement Learning!
2025
Developing a chess plugin for Manim!
2024
EazyPredict ML module
2023
Developing a scraper to get NLP data using Beautiful Soup and python!
2022
Chrome Extension To Monitor Mental Health
2021
Building a Football Score Updates API with Node.js and Web Scraping
2020From the Blog
Where I've Worked
AI Research Assistant
Massachusetts General Hospital / Harvard Medical School
Built an end-to-end deep learning pipeline for CT radiation dose estimation on ~12k DICOM scans. Achieved 12.3% error reduction vs. baseline through model iteration and reproducible evaluation for clinical stakeholders.
Lead Machine Learning Engineer
BeSuperhuman.ai
Built an LLM-driven agent system 0→1 with tool-calling orchestration, browser action generation, and validation. Reduced LLM inference and ops costs by 88% through caching layers and pipeline optimization.
Built a custom SignBERT variant for continuous sign language recognition as part of a Government of India e-governance initiative; making public services accessible to the deaf community. Trained entirely on in-house ISL video data captured in real-world conditions: low light, poor camera quality, uncontrolled backgrounds. The model had to work in the field, not just on clean benchmarks.
Full Stack Developer Intern
Agrisoft Dairy
Built a full ERP system for dairy delivery logistics - a Flutter customer app, a separate driver app with route tracking, and an admin backend to tie it together. First time shipping something real people actually used.
Got something interesting? Let's talk.
Open to research collabs, interesting engineering problems, and the occasional wild idea. I respond to most emails - no recruiters pitching Java roles, please.