Six systems, one constraint
enforced in each.

Enforces the UK GDPR Article 22 boundary as a pre-generation pipeline stage, not a disclaimer. Recommended against a GPU migration that had cleared four exit gates, after benchmarking the real cost at 28x.

Architected a three-agent retrieval system anchoring generated questions in a candidate's own material. Led a group of eight, placing third across eleven competing groups.

Cut query latency from sub-600ms to sub-100ms with hybrid BM25 and dense retrieval. Raised recall@10 by 35% over a pure semantic baseline.

Spearheaded a Neo4j knowledge graph as an LLM reasoning substrate for a global consumer brand, the project that sparked a deliberate pivot toward AI.

Engineered dual-hash, element-level change detection distinguishing model-based from property-based changes, collapsing multi-hour batch processing to sub-100ms on hundred-GB models.

Delivered real-time model-tracking technology later entering a definitive acquisition agreement with Autodesk. Developed technical capability for 30+ engineers, backing a platform onboarded by a 200+ person enterprise team.