Former industrial designer, self-taught in AI since 2020. I build the AI tools my research needs — and I measure them until they stop lying.
| 🥇 NeurIPS 2026 — RealPDE Challenge | 1st in phase 1, Track 1 (Sim2Real) · 4th, Track 2 — repo |
| 🥈 Kaggle — PTCG AI Battle Challenge | 160th / 6,807 · silver medal — repo |
| 🛢️ Kaggle — ROGII Wellbore Geology Prediction | 858th / 6,125 (top 14 %) — repo |
| 🔢 OEIS A038372 | 24 new terms, each with a machine-checkable proof — certificates |
- Proposer ≠ judge — models propose; Lean 4, SAT/DRAT or interval arithmetic decide.
- Measure, don't assert — seeded, paired evaluation; every number traces back to an artifact.
- Refutations are first-class — killed ideas are logged so no dead end is walked twice.
- Offline & private by design — local, air-gapped inference when the data deserves it.
| Project | What it is |
|---|---|
| aegaeon-showcase | Substrate of verified laws: LLM encoders, Lean 4 kernel, DRAT certificates |
| realpde-neurips-2026 | Real turbulent wake forecasting: FNO + calibrated uncertainty bounds, reverse-engineered scoring |
| pokemon-tcg-ai-battle | Transformer policy trained by imitation + seeded paired-arena evaluation |
| A038372-certificates | New OEIS terms with witness sets, DRAT proofs and Lean 4 checks |
| pam-trading-bot | Paper-trading research platform: a council of 24 archetypes, self-calibrating |
| the-internet-florist | Procedural SVG flower library, vision pipeline (SAM, Depth Anything), web app |
| harmonies | Online board-game adaptation: pure rules engine, bots, Supabase realtime, PWA |


