Christian Harley Cooper, CFA, FRM

Quantitative finance practitioner and machine learning engineer building small-model reasoning systems, language-preservation tools, and visual AI workflows.

About

I work at the intersection of quantitative finance, reinforcement learning, low-resource NLP, and mathematical visualization. My recent projects focus on training compact open models with structured rewards, building datasets for Indigenous language revitalization, and using animation/code generation as a practical benchmark for model reasoning.

My background in trading, risk, and professional finance education shapes how I build: reproducible artifacts, explicit evaluation loops, and tools that make complex systems easier to inspect.

Current Focus

Training Qwen-family models on Dakota and Stoney Nakoda language data with GRPO and grammar-aware reward functions.
Building reproducible Hugging Face model, dataset, and Space releases for low-resource NLP research.
Developing Math-To-Manim and related Manim pipelines for turning technical prompts into inspectable visual explanations.
Applying quantitative finance experience to model evaluation, risk-aware tooling, and educational AI systems.

Selected Work

Indigenous Language Preservation

I build datasets, training pipelines, and demos for Dakota and Stoney Nakoda language work, including synthetic corpora, real corpus preparation, bilingual QA data, and grammar-conditioned model training.

Representative artifacts:

Small-Model Reasoning

I experiment with reinforcement learning and reward shaping on compact models, especially for mathematical and categorical reasoning tasks such as AQuA-RAT, GSM8K-style traces, and Open-R1-style math reasoning.

Representative artifacts:

Mathematical Visualization

I maintain Math-To-Manim, an open-source system for generating mathematical and physics animations from text and image prompts. The project explores Manim generation, prerequisite discovery, multi-agent planning, and animation as a reasoning benchmark for LLMs.

Related projects include:

Math-To-Manim
KimiK2Manim
M2M2, a typed pipeline rewrite for reproducible planning, generation, review, and render artifacts.

Technical Areas

Area	Tools and Methods
Machine Learning	PyTorch, Transformers, Hugging Face Hub, GRPO, reward shaping
Low-Resource NLP	Synthetic data generation, grammar-aware rewards, bilingual corpora
Quantitative Finance	Python, derivatives, risk modeling, CFA/FRM domain knowledge
Visualization	Manim, computational geometry, 3D mathematical animation
Agentic Tooling	Multi-agent pipelines, code generation, evaluation loops
Infrastructure	GitHub Actions, Docker, Spaces, reproducible model and dataset releases

HarleyCooper
/

about-me