Selected work
Applied AI, machine learning and data engineering in delivery contexts I can stand behind, with evidence labels kept explicit. The Evidence Atlas remains the broader corpus of research, benchmarks, demos and methodology.
I prepare artefacts for readiness, model or configuration choice and regression protection, plus a subordinate interactive demo when you want to poke the behaviour.
Reproducible RAG evaluation study
Research evidence: I built this study to test when a RAG system should answer or defer using a versioned synthetic benchmark and a fixed evaluation protocol. It is not client results and not a packaged commercial offer.
Synthetic benchmark
Compares model or configuration candidates under a frozen synthetic evaluation so a team can decide whether any option clears an agreed quality bar. Not a RAG-specific product.
Observed synthetic outcome: NO ELIGIBLE RECOMMENDATION. Holdout: CONSUMED.
Synthetic benchmark only. Controlled demo, not a client result or production recommendation.
Synthetic gate demo
Shows how a release gate can block a change that introduces a critical behaviour regression: baseline PASS, one prompt mutation, critical failure, candidate FAIL, release BLOCKED.
Evidence class: SYNTHETIC REGRESSION GATE DEMO (fixture-first).
Fixture-first synthetic demo. Not live GenAI causal proof or a model-selection product.
Interactive synthetic demo
A subordinate interactive demo with its own synthetic examples. It does not replay the study benchmark, does not produce client metrics, and is not a packaged offer.
I modernise data workflows, improve quality controls and make analytical processes easier to validate, operate and hand over.
Professional data engineering delivery
I worked on the migration of a recurring analytical process from legacy SQL Server logic to a modular Databricks and PySpark workflow with data-quality and validation controls.
Public synthetic implementation
I built a local PySpark reference pipeline with explicit contracts, quarantine, reconciliation, idempotent runs, tests and CI on fully synthetic B2B wholesale data. It does not reproduce an employer's code, data or systems.