Open-source work, built to explore what production-grade AI looks like.

Multi-step LangGraph agent for loan review with conditional routing: document checks, risk scoring, compliance screening, and audit-friendly branches — modelled on real banking KYC workflows.

The unglamorous half of ML: MLflow registry, KS/chi-squared drift detection, and auto-retraining when performance degrades. FastAPI + Docker deployment.

LoRA fine-tuning of Qwen2.5-1.5B on 20k+ support pairs across 27 intents. ROUGE-1 jumped from 0.31 to 0.58, ROUGE-L from 0.24 to 0.52 — real weight updates, not prompting.

A/B testing done right, plus causal inference when randomisation isn't possible. Naive comparison estimated +29.8; DiD recovered 15.1 and PSM 14.6 vs the injected ground-truth effect of 15.0.

Fully local RAG chatbot for PDFs and URLs — LangChain, ChromaDB, MiniLM embeddings, LLaMA 3 via Ollama — with RAGAS faithfulness/relevancy scoring. No API keys required.

Probability-of-default model on Home Credit data, evaluated with industry-standard AUC and KS statistic rather than raw accuracy. LightGBM AUC 0.77 / KS 0.41 vs Logistic AUC 0.72.

ResNet18 transfer-learning classifier on 5,863 chest X-rays with Grad-CAM heatmaps for clinical interpretability. Class-weighted loss to prioritise recall — ~95% recall, AUC ~0.96.

LLM-powered structured extraction from contracts — entities, clauses, obligations — with Pydantic-enforced output schemas so downstream systems can trust the shape of what comes back.