I build verifiable training loops: pipelines that turn a single historical source — an 1890 Dakota grammar, an 1865 Cree dictionary, a 1959 railroad rulebook — into extracted rules, labeled tasks, deterministic reward functions, and published models. Plus the project's other half: Math-To-Manim ( 2,400+), a prompt-to-animation engine that turns questions into computed mathematical film.

What I do that employers actually hire for:

• Model training & fine-tuning — GRPO / RL post-training with deterministic, decomposed reward functions; LoRA adapters from 0.5B to 35B on Tinker and Prime Intellect; full run cards, reward ledgers, and audits published on Hugging Face and W&B.
• Data labeling & dataset engineering — VLM extraction from archival scans, orthography-preserving labeling, synthetic Q&A expansion, structural holdouts, hash-addressed dataset artifacts with citations intact.
• Reward/verifier design — grammar rules compiled into executable, per-component reward channels (no LLM judge; every gradient is inspectable).
• Visualization that explains — Manim render pipelines, RL training-curve dashboards, LiDAR terrain viewers.

Ask a question → get a freakin' movie. Nothing keyframed — every frame is integrated, simulated, or derived. 2,400+ stars.

The Traitor Axis Rigid-body chaos from RK4 integration — predicted flip at 3.4 s, simulated at 3.5 s.	The Last Day One continuous 3D take: sphere → torus → helicoid → catenoid → Lorenz.	Vortex Leapfrog Two rings leapfrogging, simulated live by Biot-Savart integration.
Archimedean solids	Lorenz attractor	Volatility surface — options intuition in motion

The signature move: grammar as a reward function. Take a source document, compile its rules into verifiable tasks, and let GRPO optimize against per-component reward channels — orthography, morphology, semantics — with zero LLM-judge fuzz. The reward ledger reconciles to the step.

reward = (
    0.4 * character_preservation +   # orthography: ŋ š ć ḣ preserved?
    0.4 * affix_accuracy +           # morphology: correct affixes applied?
    0.2 * semantic_correctness       # semantics: meaning vs. ground truth
) * difficulty_multiplier            # curriculum weight, 1.0x → 2.0x

Qwen3.6-35B Dakota GRPO — 82.05M tokens, composite reward climbing, ledger audit flat at zero.	Reward progression — every channel logged per step on Weights & Biases.
Railroad Engineer 1959 — a rulebook becomes a training environment.	nanochat × AQuA-RAT — small-model algebra reasoning, end-to-end RL.

volume2gym is the general compiler behind the language work: any structured volume becomes an RL gym. Sections name the world, rules constrain action, procedures encode order, exceptions define edge cases. The compiler emits cited knowledge units, six task families, grouped holdouts, deterministic reward ledgers, and SFT/GRPO trainer exports — hash-addressed and tamper-evident.

The 1959 Consolidated Code of Operating Rules lineage: 536 extracted rules → 2,708 scenarios → gym → Qwen3-4B adapter → Rule 99 contract fixture on Hugging Face.

A research line in three languages: give an endangered language one good historical book, and train. VLM extraction reads the scans — diacritics, letterpress ligatures, and all — then synthetic Q&A multiplies the surface area, then GRPO with a rubric built from the book itself. The final stage belongs to the community: speakers correct the model, and the corrections become the next training round. The model is a toddler that has read the book cover to cover; the community teaches it the rest.

Cree1865 — Watkins' 1865 dictionary, 98 pages sampled	The marks that make it Cree — diacritics as verifiable signal	Two directions — English→Cree and Cree→English
Dakota1890 — Riggs' 1890 grammar: 1,497 rules → 10,576 verifiable tasks	StoneyNakoda — Dawson's Bow Valley survey; the community-in-the-loop origin

Handwriting and OCR lineage runs through the repo list too — PyLaia (handwritten document analysis), deepseek-ocr, olmocr (PDF linearization for training data), and a reproduction of LeCun 1989 handwritten zip-code recognition — the ancestor of all of this.

Every training run ships with its curves public — reward channels, entropy, ledger audits, per-step components.

Live — refreshed every 6 hours by a GitHub Action from CNBC, Reuters, and FT feeds.

GitHub · Hugging Face · Weights & Biases · LinkedIn · X · Kaggle

Open source is the portfolio. The best entry points are the project READMEs, model cards, dataset cards, W&B runs, and demos linked above.