Projects

An end-to-end production pipeline: Azure ML training with accuracy/F1 quality gates, automatic model versioning in a registry the API pulls from on startup, GitHub Actions CI/CD, multi-target Docker deploys (Azure + on-prem via Portainer GitOps), and an enforced 90% test-coverage bar.

A multi-service system with its own data ingestion pipeline (Selenium-based scrapers + PDF ingestion into a Chroma vector store), an LLM-supervisor moderation layer, an async PostgreSQL store with Alembic migrations, a feedback admin panel, and a Dockerised multi-container deployment.

Tracked every training run in Weights & Biases — loss curves, segmentation metrics and model comparisons — and built a baseline-first pipeline that made each modelling decision reproducible and auditable.

Compared transcription engines on Word Error Rate and classical-vs-Transformer models on a common evaluation, then wired explainability (attention visualisation) into the loop to diagnose failure modes.

Modelled the full terminal lifecycle in DAX and built automated anomaly flags so operations could spot data-integrity and compliance issues without manual review.

A depth-sensing pipeline built around per-cycle baseline subtraction and blob detection, mapped through a calibration homography from sensor space to street coordinates, with a hardware-driven recalibration workflow after every physical repositioning of the camera.

Iterated from a baseline network to transfer-learning models against a human-level benchmark, and built Grad-CAM explainability plus a desktop app with confidence scores and decision logs for production auditability.

Research in progress — preliminary EDA, gap analysis against current state-of-the-art frameworks, and a defined research question driving controlled experiments.