I build AI systems that work outside a demo.

Multi-agent data analyst

A production system that runs weekly creative performance analysis across paid ad accounts without a human opening a dashboard. Designed, built and shipped solo: agent orchestration, prompt design, schema, and the delivery layer the client actually reads.

The interesting constraint was that the data model did not match reality. Ad targeting metadata existed nowhere structured; it lived inside the ad name strings, under four different naming grammars across four accounts. I reverse-engineered each grammar into a parser, which became the ingestion layer: a daily job that writes one row per live ad with rolling week-over-week metrics and a computed diagnostic status. The system is read-only on the ad accounts by design. It reports; it never spends.

A failure worth keeping

A media-fetch step in one sub-pipeline reported complete success; webhook fired, HTTP 200; while writing a 21-byte file. Nothing threw. The pipeline carried on analysing an empty asset. Diagnosing it meant not trusting the transport-level signal and adding content assertions at each hand-off. Silent success is the failure mode distributed pipelines actually die of.

Pharmacovigilance multi-agent workflow

An eight-agent workflow for pharmaceutical complaints and adverse-event intake, built in OpenAI's Agent Builder and served through a managed workflow invocation. Three routing paths separate product-information questions, commercial claims and safety-critical reports; a completeness branch checks the patient information a case needs before anything escalates to a human.

Guardrails are explicit on both ends: input filtering on what enters the workflow, output constraints on what any agent is allowed to assert. Retrieval and MCP tool actions supply the reference material. Any case that may involve a serious adverse event goes to a human regardless of how confident the model is; in pharmacovigilance patient safety outranks throughput, and the routing has to say so structurally.

Sales & procurement intelligence

Thirty supplier purchase orders spanning several years existed only as PDFs. I parsed them into a structured multi-sheet workbook; SKU normalisation, Pareto-style spend ranking, supplier-by-period comparison; and in doing so found a bonus-kit structure in the supplier's terms that produced substantial annual bonus inventory at effectively zero tax exposure. The finding was confirmed with the client's accountant.

The second half is a sales analytics layer over 275 clients: revenue analysis, dormant-account outreach targeting, and a custom SVG bubble map of accounts by geography and value. Excel is the delivery layer, deliberately; the people making the decisions were never going to open a dashboard.

What the data did wrong

• Cost and profit columns unreliable at source; recomputed rather than trusted.
• Duplicate client codes for the same real client.
• Pass-through codes that silently corrupt naive aggregation.