I build AI systems
that turn complex data
into decisions.

Problem

HESA statutory returns are a regulated workflow with hard deadlines. Missed sign-off gets reported to the Office for Students. The existing process ran on Banner extracts, Python and Alteryx scripts, Excel trackers, email chains. Hundreds of quality rules per cycle, no central view, no audit trail.

What it does

Governance and submission pipeline in one tool. Role-based access, invitations, append-only audit log, multi-institution dashboard with risk scoring. The pipeline handles XML upload, lxml-based XSD validation, OVT quality report ingestion, per-rule triage and team assignment, failure drill-down, and Core File generation from the 28 TSV outputs HESA returns after sign-off.

In progress

LLM module on Azure OpenAI with RAG over 200+ pages of regulatory guidance. Natural-language rule queries, tolerance-request drafting, schema-change summaries. Next phase: an agent layer that updates pipeline scripts behind a review gate, generates visualisations on demand, and drives the triage loop from intent.

What it is

A mobile-first personal operating system. Today view with energy tracking, calendar, goals, check-ins, analytics, and a conversational assistant with persistent memory. Full-stack, deployed, in daily use since Day 31.

Architecture decision

The memory system went through three versions in three days. Store everything. Inject everything. Score and select. The first two were thorough. Only the third was useful. The problem was never storage — it was knowing what matters right now. The assistant uses selective injection: memories are scored against the current context and only passed to the model above a relevance threshold.

What I learned

Working and right are different things. The lesson wasn't to slow down. It was to know what I was optimising for before I started.