About
Currently training language models at the Biodesign Institute and building RL environments for agents at Polymath (YC W26). On the side, I've been researching alignment, post-training and multi-agent systems.
In an agent-native world, I am obsessed with rabbit-holes of engineering and building things that people love. I love travelling and doing hackathons (recently won two awards at TreeHacks 2026). In my free time, I'm grinding chess or cycling until I run out of roads :)
I Grew Up in Six Cities
Here are some of my favorite clicks:
Work Experience
The Biodesign Institute
- Optimized GPU usage by implementing distributed training with gradient checkpointing and memory-efficient batching using JAX and Slurm, cutting training walltime from 25 to 7 hours.
- Increased model throughput by 13x and reduced parameter count by 85% with knowledge distillation while achieving predictions within 3% of the state-of-the-art teacher model.
Polymath (YC W26)
HealthGC
- Engineered a sub-200ms (P90) retrieval augmented generation (RAG) system with HIPAA compliance on multimodal data using PostgreSQL, Google Cloud, WebAssembly and JavaScript to personalize LLM responses.
- Developed tool calling functionality for web search and RAG for an LLM-based chatbot using Google Vertex AI, enabling contextual and up-to-date responses for 345k calls and 8.4 million messages by 40,000 people.
- Engineered fault-tolerant webhook infrastructure using Websockets, Express.js, and Docker achieving 99.9% uptime and 35% reduction in message delivery latency with graceful error handling.
Data Mining and Machine Learning Lab
- Built a multi-agent benchmark to probe alignment conflicts in tool-calling LLM agents. Ran it on a Slurm managed HPC cluster, orchestrating 12 agent configurations across 9 models (650+ GPU-hours).
RCV Innovations
- Designed Power BI dashboards for A/B Testing which identified 3 critical user experience bottlenecks.
Education
Faking a Virtual Filesystem Over GitHub's API for Browser-based Agents
Engineering details on how I built git-fs, a coding agent that runs completely inside your browser.
Should MCP Servers be Agents in Disguise?
Blog post + Case study on why 'under-the-hood' agents can improve the token efficiency of MCP servers.
How Computation Graphs solve the Eager Execution Problem in PyTorch (torch.compile)
Technical deep-dive into why computation graphs help with memory management and compute amortization in PyTorch.
Check out my latest work
I've worked on a variety of projects, from websites to scalable backend systems for AI agents. Here are a few of my favorites.
Faking a filesystem over GitHub API to run coding agents and instantly chat with any repository fully in your browser. Wrote a blog about this.
Canvas MCP
MCP Server that allows you to query your courses in the AI app of your choice. Over 2K tool calls in April 2026.
Polymarket MCP
Lets agents access Polymarket market data in the AI app of your choice. Over 50K tool calls on Smithery's leaderboard of April 2026.
Turing
Turing watches how you do tasks, how you reply to messages, how you work to learn your patterns and automate your tasks.
irina (AI for Minesweeper)
AI that solves Minesweeper in record-breaking time. Uses a probabilistic approach, scans the grid, trains a Reinforcement Learning model to make decisions.
SEOthing
CLI tool to skyrocket your Lighthouse score by converting images to WebP format, generating meta tags, and optimizing for better performance. npx seothing
I like hacking things (a lot)
During my time in university, I attended 11+ hackathons. It was eye-opening to see the endless possibilities brought to life by a group of motivated and passionate individuals within 2-3 days. I have made some of my best friends and memories at these hackathons :)
- T
TreeHacks 2026 (2x Sponsor Track Winner 🏆)
Jiggle WiggleStanford, CAJiggle Wiggle is a real-time AI movement coach that uses motion capture to analyze your body's movements and provide feedback to improve your limb coordination and balance. We won the Zoom x Render Track and runner-up award in Modal's AI Inference Track. - C
CalHacks 2025
TuringBerkeley, CATuring watches how you do tasks, how you reply to messages, how you work, and personalizes your computer-use agent so your workflows can be automated based on your patterns and habits. - H
HackMIT 2025
FlashbackCambridge, MAWe built a RAG for videos and developed it with smart glasses as the base to create a social media platform. I worked on facial recognition and task matching algorithm using OpenCV and Modal's Compute Platform, so the RAG could match the user's contacts to the video content. - H
HackIllinois 2025 (Microgrant 🏆)
AliceUrbana-Champaign, ILBuilt a chrome extension allowing researchers to learn more about the papers cited by simply hovering over the citations. Learned a ton about Model Context Protocol (MCPs) and PDF.js rendering engine. - C
CalHacks 2024
CuraRxBerkeley, CABuilt a full stack web-app that allows users to track their medication schedules. Designed a multi-modal Agentic AI System to check-in on patients and detect for potential side effects using sentiment analysis. - H
HackMIT 2024 (Sponsor Prize 🏆)
InsightHireCambridge, MAAwarded 1000$ in `Best Use of AI without an API`! Built a Python-Flask backend that uses Language Models to re-rank resumes and web-crawl to validate applicant resumes. - S
- L
LAHacks 2024
FreshUntilLos Angeles, CABuilt a full stack web-app designed to help individuals manage their groceries more efficiently. We aim to Minimize Waste & Maximize Taste. - W
WiCS Coding Contest 2024 + HKD Code-a-thon 2024
Tempe, AZWon 1st Place Undergraduate Team (out of 150+ participants) at WiCS and 1st Place Overall at HKD Code-a-thon. Both of these were Leetcode-style coding competitions with questions in Data Structures and Algorithms. - A
AI in Education by ScaleU (Honorable Mention 🏆)
Scottsdale, AZ - O
Opportunity Hack 2023 (Finalist)
Tempe, AZTop 5 Finalist out of 100+ participants
Skills & Tech Stack
Get in Touch
Want to chat? Just shoot me a text on twitter or email me directly at contact@aryankeluskar.com.