Ayush Pai

I like to work on interesting problems in ML and systems. I just graduated from Georgia Tech and am moving to San Francisco very soon. If you're looking for roommates, reach out :)

In my free time, I like photography, tinkering, and kudo-farming on Strava.


Current

Transition - Traveling, tinkering, moving to SF.

Researcher @ EI-HCAI Lab, Georgia Tech - Working on novel guidance techniques for generative infilling in diffusion models.


Past

Georgia Tech - Graduated with a bachelor's and master's in Computer Science.

Second Thoughts - A hybrid decoding pipeline using GPT-2 and MDLM that masks low-confidence autoregressive tokens and regenerates them with bidirectional diffusion context, improving on HellaSwag, WikiText-103, and MAUVE.

ML Intern @ Meta - Built a ranking model for Instagram that decides how content is ordered in production.

Donna - A local file search system that turns files into vectors, rewrites a query into related searches, and retrieves the closest matches by meaning instead of exact text, with 49% better retrieval accuracy than Spotlight (macOS Sequoia, 2024).

SWE Intern @ AWS - Built a Lambda-based validation system for AWS Organizations that cut validation runtime from 5 hours to under 2 minutes.

Open Source Drawing Calculator - An open-source version of Apple’s iOS 18 Math Notes that reads hand-drawn equations, parses them with a multimodal model, and writes the computed answers back onto the canvas.

Sideline - An AI referee system for basketball that recognizes fouls from game footage; pitched to NBA executives and featured on Overtime Sports.

Rentathon - Built a person-to-person rental marketplace for household items.