Things I build
Robots, mechanisms and parts I design and print for my own problems.
Write-ups comingAbout
Hi, I'm Varshil.
I'm curious by nature and a builder at heart. I rarely say no to trying something new, and once I start, I go all in. Science and engineering have been part of me for as long as I can remember, and I still build things mostly for the joy of it.
At work, I lead AI at AirMDR, where we're building an AI SOC: agents that run cybersecurity operations on their own. Before that, I spent almost four years at Microsoft training and deploying the language models behind its Ads recommendation system.
At IIT Guwahati, I mostly tinkered with robots and reinforcement learning. That led to internships at Intel Labs, IIT Madras and MILA, and a seat in Carnegie Mellon's master's programme, which I gave up when COVID hit.
Robots were my first love, which is why lately I've been working on some ambitious ideas in Physical AI.
Outside work, you'll find me gardening, 3D printing something or looking up at the night sky. Sometimes I'm skating or bird watching in Cubbon Park.
Experience
-
2024now
Leading AI for an agentic SOC that detects, investigates and responds to security alerts on its own.
- Built the agent harness: model-agnostic, 150+ integrations, onboarding in hours.
- Launched FAST, our self-serve AI SOC, plus a security copilot and playbook builder.
-
20202024
Microsoft
Trained language models for a living, optimising them for Microsoft's ad recommendation system.
- Data & Applied Scientist 22023 – 2024
Led short-text relevance for the Ads stack. In-house LLM labelling saved $500K a year.
- Data & Applied Scientist2020 – 2023
Took relevance models to 70+ languages, at a third of the compute.
- Data & Applied Scientist 22023 – 2024
-
20162020
IIT Guwahati
Mostly built robots and studied reinforcement learning, and spent my summers at some very cool places.
RISE Lab, IIT Madras2019Graph-based priors that help autonomous driving generalise to new roads.
MILA2019Transformer language models that help RL agents play text games.
Intel Labs2018Imitation learning (GAIL) that teaches robot arms from expert demos.
Projects
CADForge
An RL environment where agents learn to write CadQuery code that becomes manufacture-ready STEP parts, rewarded by pure geometry.
Read the write-up →
Kaggle Wins
A gold in American Express Default Prediction (16th) and a silver in Feedback Prize (14th), and what each solution did.
Read the write-up →
Earlier work
Placement experience: Microsoft Data Science
Advice for anyone preparing for data scientist roles.
Learned representations for speech
Bachelor thesis: self-supervised pre-training for Indic code-switched ASR.
AnimalAI
PPO with behaviour cloning for the Animal-AI Olympics.
BERT on NarrativeQA
Reading comprehension over long stories.
ARLE
A robot that picks and places books in the IIT Guwahati library.
Ship detection device
A low-cost collision alarm for fishing boats. Bronze at the Inter-IIT Tech Meet.
Robotic palm
Hand-pose estimation driving a robotic hand from live video.
Linear inverse RL
Linear inverse RL algorithms on Mountain Car.
Lunar Lander
Policy-gradient algorithms on Lunar Lander.
Life
Not work, just the stuff I've done here and there.
Orion Nebula, March 2021Astrophotography
A Celestron AstroMaster 130EQ and a phone held to the eyepiece.
Dragon fruit flower, November 2021Gardening
A garden grown entirely in pots. Vegetables first, then flowers, then fruit.
Roller skating
Raced for years, and won at the national level in 2013.