Open-Source Projects for CAPIX AI / Large Quantitative Model Development

A curated set of open-source projects worth evaluating for CAPIX’s AI and large quantitative model work, organized by role in the stack and focused on treasury / capital-markets quantitative modeling.

Internal planning document. Not published to the website.

1. Core ML / model-training frameworks

The foundation layer for training or fine-tuning large models.

  • PyTorch — de facto standard for research and production model development. Mature ecosystem, best hiring pool.
  • JAX + Flax — Google’s high-performance autodiff/XLA stack. Strong for large-scale numerical and quantitative work where you want compiled, vectorized math (closer to “quant” code than typical deep learning).
  • Hugging Face Transformers / PEFT / Accelerate — the practical toolkit for fine-tuning and deploying transformer models (including LoRA/QLoRA parameter-efficient tuning, which keeps GPU costs sane).
  • Ray — distributed compute for scaling training, hyperparameter tuning (Ray Tune), and serving across a cluster.

2. Time-series & financial quantitative modeling

The most directly relevant category for treasury/markets.

  • Qlib (Microsoft) — purpose-built AI quant investment platform: data handling, factor models, backtesting, and ML model pipelines for financial forecasting. Probably the single most relevant project to the CAPIX domain.
  • GluonTS (Amazon) and NeuralForecast / StatsForecast (Nixtla) — probabilistic and neural time-series forecasting (cash-flow, FX, rate, liquidity forecasting).
  • Darts — unified API spanning classical (ARIMA) to deep-learning forecasters; good for prototyping treasury forecasting models.
  • QuantLib — the established library for derivative pricing, yield curves, fixed income, and risk — useful as the “ground truth” quantitative engine that ML models augment rather than replace.
  • Prophet — quick, interpretable baseline forecasting.

3. LLM / foundation-model serving & orchestration

If “large quantitative model” includes LLM-augmented analytics or agentic workflows.

  • vLLM — high-throughput LLM inference/serving (the standard for self-hosting open-weight models efficiently).
  • Open-weight model families — Llama, Mistral / Mixtral, Qwen, DeepSeek — for self-hosted models where data residency/confidentiality of treasury data matters.
  • LangChain / LlamaIndex — orchestration and RAG over internal documents/data.
  • Ollama — easy local model hosting for dev/experimentation.

4. Experiment tracking, data & pipeline infrastructure

  • MLflow — experiment tracking, model registry, lifecycle management.
  • DVC — data and model versioning (critical for reproducibility and audit — important in a regulated financial context).
  • Feast — feature store for serving consistent features to training and inference.
  • Apache Arrow / Polars / DuckDB — fast columnar data processing for large market/time-series datasets (Polars and DuckDB in particular are excellent for quant data wrangling).
  • Optuna — hyperparameter optimization.

5. Validation, explainability & risk

Matters for finance and regulatory governance.

  • SHAP / Captum — model explainability, important for governance and regulator/auditor scrutiny.
  • Evidently — data and model drift monitoring in production.
  • Great Expectations — data quality/validation gates in pipelines.

A pragmatic starting stack for CAPIX

A minimal, coherent combination to begin with:

PyTorch + Qlib + Polars/DuckDB + MLflow + DVC, adding vLLM + an open-weight model if self-hosted LLM capability over confidential treasury data is needed.

Decisions to make before committing

These change the recommendation materially:

  1. What is the “large quantitative model” — a forecasting/risk model (favors Qlib, GluonTS, QuantLib), or an LLM-based analytics/agent system (favors vLLM, open-weight LLMs, LangChain)?
  2. Self-hosted vs. API — does treasury-data confidentiality require everything on-prem/VPC, or can hosted model APIs be used for some parts?
  3. License constraints — most above are permissive (Apache/MIT/BSD); a few model weights have usage restrictions worth checking against commercial use.