DaveA'Hearne

End-to-end MLOps pipeline predicting the probability of each driver finishing on the podium for a given F1 race. Trained on historical Ergast data from 1990 to 2024, validated against live 2025 results, and served via a containerised FastAPI application. Documented as a multi-part blog series covering the full arc from EDA and heuristic baselines through to automated retraining and drift-gated model promotion.

• LightGBM binary classifier trained with walk-forward cross-validation across 10-year rolling windows, with engineered features covering driver and constructor rolling podium rates, championship position, circuit-specific form, regulation era, and mechanical DNF rates.
• LakeFS provides Git-style data versioning with a staging-to-main branch strategy; every training run links to the exact LakeFS commit SHA via MLflow dataset logging, giving full data-to-model lineage.
• Model exported to ONNX at training time and registered in the MLflow model registry; deployment is an alias swap — the serve layer fetches by alias so promotion requires no code change and no redeploy.
• Evidently drift detection runs after each race weekend against a pinned reference dataset; a Prefect flow orchestrates ingestion, validation, conditional retraining, and margin-gated promotion end to end.

Clinical code RAG pipeline built for NICE (National Institute for Health and Care Excellence), developed as the employer project for the Cambridge Data Science Career Accelerator. Given a free-text clinical research question, the system parses it into typed clinical entities, retrieves and ranks relevant codes across SNOMED CT, ICD-10, QOF, and NHS reference sets, and returns results with source attribution and confidence scoring.

• Two-stage RAG pipeline: offline ingestion encoding clinical code descriptions into biomedical embeddings (SapBERT/BioBERT) stored in sqlite-vec, and an online query stage combining semantic vector search with TF-IDF hybrid retrieval, reranking, and LLM-assisted reasoning.
• Stanza NER (bc5cdr and i2b2 biomedical models) decomposes free-text queries into typed entities (conditions, drugs, exclusions), driving structured retrieval across multiple code systems simultaneously.
• FastAPI service with WebSocket streaming of pipeline progress so results appear incrementally as each stage completes.
• MLflow evaluation tracking F1, precision, and recall against gold-standard question-to-code-set pairs, with an acceptance threshold of 0.70 across all three metrics and confidence scoring above 0.65.
• Dockerised with automatic ingestion on first boot and Stanza model caching baked into the image layer.

End-to-end image classification pipeline built as an MLOps exercise. A PyTorch CNN trained on CIFAR-10 and exported to ONNX for runtime decoupling, served via a production-style FastAPI REST API. Emphasis on the gap between a working model and a deployable one.

• PyTorch CNN with a 90/10 train/validation split, per-epoch checkpointing, and a held-out test set evaluated only at the end — ONNX exported with a pinned opset version to prevent silent runtime drift.
• FastAPI serving layer with API key auth, request ID propagation through context vars, structured log formatting, rotating file handler, and a timeout middleware returning 504 on breach.
• Separate Docker images for training and serving: the train image persists checkpoints and the ONNX artifact via mounted volumes; the serve image mounts the ONNX file at runtime rather than baking it in.
• Test suite using pytest with a mocked ONNX InferenceSession, covering preprocessing shape contracts, softmax numerical stability, postprocessing label mapping, and the inference endpoint.

Four analytical projects completed across the Cambridge Data Science Career Accelerator (PACE), covering unsupervised learning, supervised classification, NLP, and time series forecasting. Each applied to a real-world dataset with a defined business problem and written report.

Two-sided senior tech recruitment platform for the UK market, built and run as a founder project. Hard constraint filtering into BAAI/bge-large-en-v1.5 sentence transformer scoring for candidate-role matching. A Scout service pulls from ATS adapters across Greenhouse, Lever, Workable, and Ashby. Go chosen deliberately for compiled binaries, minimal Alpine container images, and first-class concurrency via channels wired into WebSocket connections.