Production ML systems, compliance frameworks, and data science case studies — grouped by domain.
SR 11-7 & EU AI Act compliant validation suite for financial LLMs
A local-first auditing suite that automates Model Risk Management artifacts regulators require — converting raw LLM outputs into compliant evidence. Features adversarial red-teaming (50+ templates), SHAP/LIME token attribution, Financial-F1 accuracy scoring, and a YAML-based regulatory mapping registry.
Sales forecast pipeline + GenAI narrative generation
End-to-end time-series forecasting that predicts sales trends, then uses an LLM (via meerax OllamaProvider) to write plain-English summaries — bridging analyst output and executive communication. v0.1.5, meerax doctor clean, zero open PRs.
Four recommenders + a hybrid blend, explained in plain English
User-CF, item-CF, content-based TF-IDF, and ALS behind one interface with a configurable weighted hybrid, plus a GenAI layer that explains why each article was recommended. Real Precision@K/MAP@K evaluation, self-contained HTML report. v0.1.2, built entirely via subagent-driven development.
Credit-card fraud detection with explanations attached
Credit-card fraud classification paired with GenAI-written explanations for each flagged transaction, plus an AI-governance layer documenting intended use, fairness limitations, and audit trail. v0.1.1, built on meerax, shipped clean.
TensorFlow CycleGAN — unpaired image-to-image style transfer
CycleGAN retrofit on meerax adapting tensorflow_examples' hardcoded 3-channel architecture to arbitrary domains via a channel-tiling wrapper. Apple→Orange as the reference pair; extensible to any two-domain task. Full meerax project structure, CI, meerax doctor clean.
Image captioning that speaks — attention, GenAI, and real TTS
InceptionV3 → Bahdanau attention → GRU captioning trained on Flickr8k, with an LLM expanding each caption and real text-to-speech reading it aloud. The real training run surfaced and fixed two bugs no code review had caught, and a genuine evaluation-split data leak — all honestly re-reported. v0.2.0.
97% accuracy malaria classifier — FastAI transfer learning
Binary image classifier distinguishing parasitised from uninfected blood cells using FastAI's transfer learning pipeline on a pretrained ResNet. Achieves 97% accuracy on the NIH cell image dataset with minimal labelled data.
Multi-engine OCR benchmark with CER, F1, and audit trail reports
Production-grade OCR framework around PaddleOCR with F1 Score evaluation, Character Error Rate via Levenshtein distance, and timestamped audit trail reports. Replaces black-box OCR with auditable evidence — critical in regulated document intelligence.
Production ML app — Ames Iowa + Delhi NCR price prediction
Production real estate prediction portal: a log-transformed LassoCV model for Ames, Iowa with SHAP attributions and MRM diagnostics, and per-region Gradient Boosting for Delhi NCR (Gurgaon, Noida, Delhi) with a Folium map and Buy/Rent toggle.
EDA on 2007–2011 lending data to identify default drivers
Exploratory data analysis of a real lending portfolio — amount-to-income ratio, revolving utilisation, derogatory records, and loan purpose as key default signals. Univariate, bivariate, and trivariate analysis across 14.16% overall default rate; Spain flagged at 18.31%.
Ridge, Lasso, and SR 11-7 MRM diagnostics on Ames Iowa housing
End-to-end regression study on 2,919 Ames Iowa sales (80 features). Ridge vs Lasso comparison, SHAP attributions, and MRM diagnostics — VIF, Durbin-Watson 2.03, Breusch-Pagan, Jarque-Bera — aligned with SR 11-7.
In-house ML toolkit powering every project on this page
Installable Python package on PyPI. Claude/OpenAI/Ollama provider abstractions behind one interface, classification + timeseries + text eval metrics, dark-themed matplotlib plots, stratified/time splits with SMOTE, self-contained HTML report builder, and a CLI (meerax new, doctor, bump) that scaffolds and audits projects against a shared governance checklist. See the full ecosystem and live version on its own page.