The technical workshop behind cmlpolymath.dev
Six production-minded projects spanning econometrics, ML engineering, MLOps, AI governance, RAG architecture, and systems engineering in Rust. Each project is scoped around a distinct professional signal — not isolated notebook experiments, but systems built with statistical rigor, deployment constraints, and real stakeholders in mind.
| Project | Domain | Stack |
|---|---|---|
claw-voyant |
Regression pipeline in Rust | Rust · Cargo · CSV |
ez-abs |
Experimentation & A/B testing | Python · Statistics · CUPED |
governance-suite |
AI fairness & compliance | Fairlearn · SHAP · XGBoost |
macro-forcaster |
Macroeconomic forecasting | Python · ARIMA · VAR · Marimo |
ml-guardrail-sys |
MLOps guardrails & promotion | MLflow · Streamlit · Drift detection |
stag |
Enterprise RAG system | Qdrant · FastAPI · Streamlit · Local LLM |
Regression pipeline in Rust
A dataset-agnostic ML regression pipeline built entirely in Rust. Covers synthetic data generation, leakage-safe preprocessing via fit/transform type enforcement, model training, evaluation, prediction, and Markdown report generation — all from a CLI. Swap the CSV and target column; retrain.
Signal: Production data tooling outside the Python comfort zone, with attention to performance, reproducibility, and system design.
Strategic experimentation platform
A production-style A/B testing framework built around better decisions, not just p-values. Includes CUPED variance reduction, SPRT sequential monitoring with early stopping, typed KPI hierarchy design, guardrail metrics, and a written decision policy for interpreting experiment outcomes.
Signal: Experimentation systems that connect statistical inference directly to business decision-making.
AI governance and fairness compliance
End-to-end ML governance tooling: fairness auditing, bias mitigation (Fairlearn ExponentiatedGradient + ThresholdOptimizer), SHAP explainability analysis, and a generated compliance report. Covers EEOC, NYC Local Law 144, and EU AI Act frameworks with a Plotly Dash review dashboard.
Signal: ML systems that are auditable, explainable, and aligned with responsible AI practices.
Macroeconomic forecasting
A Stata-to-Python translation of econometric forecasting workflows. ADF/KPSS/Johansen stationarity testing, Auto-ARIMA, ETS, Prophet, rolling OLS with Newey-West HAC standard errors, VAR/IRF analysis, walk-forward backtesting, and Diebold-Mariano forecast comparison — with a Marimo interactive interface throughout.
Signal: Bridging legacy econometric methodology and modern Python tooling without sacrificing statistical rigor.
MLOps guardrail system
A full model lifecycle management system: training, registration, PSI/KS statistical drift detection, automated retraining triggers, champion-challenger promotion gates with five quality checks, and a Streamlit dashboard for human review. GitHub Actions retraining workflow with concurrency locking and a break-glass force-promote option.
Signal: Production ML is not just model training — it is monitoring, governance, controlled promotion, and operational continuity.
Enterprise RAG system (S-Tier rAG)
A Retrieval-Augmented Generation system built around retrieval quality and grounding. Document ingestion (PDF, EPUB, plain text), chunking, sentence-transformer embeddings, Qdrant hybrid vector search, cross-encoder reranking with logit normalisation, query rewriting, source-grounded answers, and local LLM integration via Ollama. FastAPI backend, Streamlit UI, CUDA-accelerated on Linux/WSL.
Signal: RAG systems that prioritise retrieval quality, answer grounding, observability, and practical deployment — not just demo-grade pipelines.
- Statistical discipline — inference, backtesting, stationarity, variance reduction, guardrail metrics
- Operational realism — monitoring, promotion gates, dashboards, and deployable interfaces
- Reproducibility — deterministic workflows, documented assumptions, clear configuration
- Interpretability — explainable models, readable reports, decision frameworks non-technical stakeholders can act on
- Full-path ownership — raw data → pipeline → model → reporting → deployment → monitoring
git clone https://github.com/cmlpolymath/repertoire.git
cd repertoire/<project-name>Python projects include requirements.txt or pyproject.toml. The Rust project uses Cargo. Each folder contains a project-level README with the run path.
If the website is the front door, this repository is the workshop.