Things I'm building

I learn by building.

A running list of side projects and research. Some are shipped and pip-installable, some are weekend experiments, some are papers. The throughline: multimodal models, computer vision, and football showing up more often than it probably should.

Shipped

4

Out in the world, install or play with it today.

artifold screenshot

A local-first library for the stuff you make with AI. Index, search, preview, and share your work, then use your past output as the style guide for the next thing you build.

Published on PyPI, pip-installable

  • Python
  • CLI
  • embeddings
  • semantic search
rl-arcade screenshot

A gamified app that teaches reinforcement learning across six levels of PufferLib games, with interactive in-browser training, concept visualizations, and an XP system.

6 playable levels, train models live in the browser

  • Streamlit
  • RL
  • PufferLib
Tavern Tactics screenshot

Poker meets Gwent in a medieval tavern: build poker hands across two rows, manage a 10-card budget, and outlast the house across a best-of-3 match.

Single-player strategy card game, playable in the browser

  • TypeScript
  • game
  • strategy

Research

3

Papers and applied research, from satellite imagery to ads ranking.

The Text on the Creative

repo coming soon

First-author paper on an under-exploited ranking modality for short-form video ads: the text rendered on the creative itself.

RecSys '26 Industry Track, under review

  • multimodal
  • ranking
  • OCR

Hurricane Damage Estimation

repo coming soon

Property-damage prediction from pre- and post-disaster satellite imagery paired with insurance data, using a Siamese change-detection network.

0.804 IoU, invited talk at GeoPython 2020

  • satellite imagery
  • Siamese networks
  • change detection

Building Detection in Dense Regions

repo coming soon

Co-authored work at Reliance Jio on extracting building footprints at scale to automate population estimation for the JioFiber rollout.

Reverse Mask R-CNN lifted mIoU from 75% to 84.3%

  • Mask R-CNN
  • remote sensing
  • segmentation

In progress

4

Active experiments. Some repos are still being cleaned up to share.

MLE Interview 2.0 screenshot

MLE Interview 2.0

repo coming soon

An interactive prep portal covering all nine rounds of the modern ML engineer interview: in-browser Python coding exercises via Pyodide, quizzes, and SM-2 spaced-repetition flashcards.

9 rounds, 46 topics, 16 coding exercises, 35 flashcards

  • Next.js
  • Pyodide
  • spaced repetition
AURA FC screenshot

AURA FC

repo coming soon

A real-time pipeline that turns soccer footage into live AI commentary. YOLOv8 detection and tracking feed an event layer, which drives generated play-by-play overlays.

SAHI slicing 2x'd ball recall; camera-motion compensation halved false sprints

  • YOLOv8
  • SAHI
  • computer vision
  • LLM
  • TTS

visionbench

repo coming soon

A growing set of from-scratch ML coding challenges for vision models, VLMs, and VLAs. The kind of problems worth keeping sharp on.

Spans convolution, IoU, patch embeddings, VAE reparameterization, and more

  • PyTorch
  • VLMs
  • VLAs
  • evals
graph quest screenshot

graph quest

repo coming soon

Learn graph theory and combinatorics through interactive, gamified lessons: modules from fundamentals to BFS/DFS, with XP, levels, and a graph playground on the way.

Interactive lessons with a node-and-edge visualizer

  • Next.js
  • Cytoscape
  • education