I'm Christian H. Cooper, an ML engineer working on low-resource language models, reinforcement learning, and agent harnesses. My main research line is Dakota and Cree: turning historical grammars and dictionaries into extracted data, verifiable tasks, reward functions, and models that can be evaluated and corrected.

I also build Math-To-Manim, where a mathematical question becomes a visual explanation, and WorkspaceAlberta Harness, which connects AI tools to practical work for Canadian businesses and skilled trades. Across these projects, I work on the whole loop: prepare the data, build the system, inspect its failures, and measure what improves.

The engineering behind the projects:

• Model training & fine-tuning — GRPO post-training and LoRA experiments on Tinker and Prime Intellect, with model cards, checkpoints, and training logs 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-derived checks with separate orthography, morphology, and reference-matching scores, so a change in reward can be traced to the scoring code.
• Visualization that explains — Manim render pipelines, RL training-curve dashboards, LiDAR terrain viewers.

Ask a question → get a visual explanation. Math-To-Manim works backward through prerequisites, builds a teaching sequence, checks the mathematics, and turns it into a Manim scene for rendering and review.

I started it on the morning of January 20, 2025, the day DeepSeek-R1 was released. The first project commit records 04:24 Mountain Time. GRPO made me think about recursive self-reasoning: could a system check intermediate results and use that feedback to improve its next attempt? Math-To-Manim gave me something concrete to test. R1's release record marks the same day.

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

Still in progress: I am using Math-To-Manim as an RL experiment, turning failed scenes and repair attempts into training tasks. Cheap code checks guide rollouts; rendered frames provide slower evaluation. Inference-time revision and RL weight updates are separate steps, and I am still working on the connection between them. Read the experiment notes.

Can a historical language volume provide enough structure to start a useful training loop?

Dakota1890 starts with Riggs' 1890 grammar and dictionary: preserve the orthography, extract rules, generate tasks, and score outputs against explicit constraints. The experiments have progressed from a 0.6B model to a 35B GRPO adapter.

Cree1865 tests the approach on Watkins' 1865 dictionary, with synthetic bilingual Q&A, a Cree-specific verifier, and LoRA training through Tinker. Its 800-step run, model card, and inference demo are linked below.

The shared idea is grammar as a reward function. I log the scoring components separately to see what the model is learning and where the verifier falls short. A higher score on source-derived tasks does not establish fluency; speaker-led evaluation and correction remain the next stage I want to develop.

Now building: Baguettotron-Dakota1890 connects the Dakota1890 morphology gym to a GRPO fine-tuning stack for PleIAs/Baguettotron.

One Dakota reward formulation illustrates the component breakdown; the Cree verifier uses its own rubric.

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 develops the broader engineering idea: turn structured source material into cited knowledge units, training tasks, and executable checks. It supports six task families, grouped holdouts, reward ledgers, and SFT/GRPO exports, with hashes that make changes to artifacts detectable.

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.

The scans are part of the engineering problem. Diacritics, variant spellings, and dictionary direction matter when text becomes training data. I use VLM extraction and synthetic Q&A to create more tasks from the relationships recorded in each source, while keeping the distinction between extracted material and generated examples.

These historical books are starting points. They cannot capture a living language on their own. The correction loop I want to build retains the prompt, the model's answer, and a speaker's correction with the context that explains the mistake.

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

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.

I design WorkspaceAlberta Harness, a custom AI terminal that connects business context with procurement evidence from CanadaBuys and Alberta Purchasing Connection. The aim is practical: help an operator assess an opportunity and produce a useful bid brief, document, or working tool.

My work covers connectors, tool routing, persistent task state, and agent procedures that can be improved through reviewed corrections. It applies the same discipline as the training work: keep the evidence visible, distinguish missing information from a confirmed answer, and measure whether the result helps the person doing the job.

The linked training runs expose reward curves and component metrics. I use these logs to investigate failures and check whether the recorded rewards agree with the verifier.

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

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

Interested in low-resource language modeling, RL environments, or agent harnesses? The repos and run cards above show how I work. Get in touch if you are building in the same direction.