Hey, I'm Ayush. I build AI products for healthcare, research, sports analytics, and real-time decision systems. Here's some more about me:
- Right now I'm building WorldBench — an open-source, CI-grade regression-testing tool for robotics world-model checkpoints. It's on PyPI (
pip install worldbench), ships a reusable GitHub Action for checkpoint gating, and lives at worldbench.xyz. - I love healthcare AI that does real work: MedEase (senior-friendly medication assistant, live app), Adaptive AI PT (webcam rehab coach with pose tracking, live app), and SDoHv4 (multilingual social-determinants-of-health questionnaire with voice I/O).
- I also build sports analytics tools, like a college football tendency analyzer for play-by-play data, situational success rates, and coaching-style insights.
- My favorite projects are ones where the AI or data layer has to earn its keep: analyze uncertainty, adapt to user input, explain its decisions, or make a live workflow easier — not just isolated model demos.
- I prototype fast — most of the projects below went from idea to deployed in days, not months.
- Reach me via my portfolio, Twitter/X, or email at iamthehakunamatata@gmail.com.
Languages
Front‑end
Back‑end
Data & ML
DevOps & Infra
WorldBench — regression testing for robotics world models (2026 – Present)
- CI-grade checkpoint regression testing for robotics world models — catch quality regressions before they ship.
- Published on PyPI as
worldbenchwith OIDC trusted publishing. - Ships a reusable composite GitHub Action for checkpoint gating directly in CI.
- Docs and demos at worldbench.xyz, including honest-limits documentation about which metrics activate on real robot video vs. synthetic tracking suites.
- Flagship validation: NanoWM-B/2 50k vs. 300k checkpoints across RT-1 episodes.
| Project | Tech Stack | What It Does |
|---|---|---|
| WorldBench | CI-grade regression testing for robotics world-model checkpoints. On PyPI, with a reusable Action for checkpoint gating. | |
| MedEase (live) | Senior-friendly medication assistant with auth, prescription workflows, camera verification, and caregiver alerts. | |
| Adaptive AI PT (live) | Webcam-based AI rehab coach: Claude-generated plans, pose tracking, voice cues, and adaptive form feedback. | |
| TruthLayer | Real-time AI context for news: bias scoring, narrative comparison, voice summaries, and topic memory. | |
| ColliderGuard | CERN HIGGS anomaly-detection dashboard: three ML models, PCA/t-SNE maps, and interpretable explanations. | |
| OpsTwin AI | Warehouse digital-twin simulation with autonomous robot orchestration, WebSockets, and pathfinding. | |
| SDoHv4 | Multilingual social-determinants-of-health questionnaire with voice input/output and hospital-specific exports. | |
| Football Tendency Analyzer | College football play-by-play analysis: situational success rates, run/pass tendencies, coaching-style insights. | |
| Firecracker Live Migration | Live migration experiments for Firecracker microVMs. |

