AI/ML engineering and technical leadership
I architect production AI systems end to end.
Then I measure them honestly, even when that kills the brief.
Asked to build a model router, I measured the fleet first and found that model choice was about one percent of the bill. The other 62.9% was cache writes. The recommendation I delivered was smaller than the one I was asked for.
Selected work
9 systems, in depth
Each one is written the way I would talk through it: the problem, the constraints, the architecture, the decisions I would defend, and the things that broke.
Cost architecture study · production LLM economics
What a Production Agent Fleet Actually Costs
Asked to build a per-request model router, I measured the fleet first and found model choice was about 1% of the bill.
Read→62.9%observedCache writes as a share of spend
Evaluation infrastructure · claim validity as the product
AI Benchmarking & Measurement Platform
A benchmarking platform where getting the number is the easy half. The hard half is deciding whether it is allowed to mean anything.
Read→25codeValidity gates
Multi-agent research · sanctions, adverse media, ownership
AI Compliance Investigation Platform
An analyst names a party; six sanctions regimes, adverse media, aliases and corporate ownership come back as a cited, defensible risk determination.
Read→6codeSanctions regimes screened
32 domain agents · SAP HANA · three surfaces
Enterprise NL-to-SQL Agent Platform
Planners ask supply-chain questions in plain English; 32 domain agents answer with SQL-grounded results, under role-scoped access the model cannot talk its way past.
Read→32codeDomain agents in production
Hybrid ensemble · Airflow · Snowflake · a gold set that lied
Freight Pricing Anomaly Detection
Thousands of daily freight transactions scored for over- and undercharge by three deliberately asymmetric detectors, and an evaluation that turned out to be measuring the review interface rather than the model.
Read→90%attestedReduction in manual pricing review hours
Earlier systems, reconstructed
Designed and shipped before the work above. Rebuilt here from the résumé, and labelled as such throughout.
Financial crime · 10M+ transactions · rules the regulator can read
reconstructedAML Transaction Monitoring & Case Platform
A hybrid surveillance system where a sequence model ranks alerts and a rules engine still decides what an alert is, because in anti-money-laundering, the part you cannot explain to an examiner is the part you cannot ship.
Read→10M+attestedTransactions monitored
Auto finance · delinquency and charge-off · a label that changed meaning mid-year
reconstructedCredit Risk Early Warning & Collections Prioritisation
Models that predicted which accounts would roll to charge-off early enough to do something about it, built just as a payment-holiday programme made the word 'delinquent' mean something different.
Read→87%attestedModel accuracy
Document extraction · millions of freight records · every discrepancy, not a sample
reconstructedFreight Invoice Reconciliation at Document Scale
Carrier invoices arrive as PDFs and the system of record holds what the shipment should have cost. Reconciling the two across millions of records turned an audit that had always been a sample into one that was complete.
Read→100%attestedDiscrepancy identification
Asset recovery · the first model I put in front of a business decision
reconstructedVehicle Recovery Optimisation & Reporting Automation
Repossessed vehicles lose value every day they sit. A model that ranked recovery routes by expected net proceeds improved recovery profits 30%, and the reporting automation around it is where I learned what makes an analysis get used.
Read→+30%attestedVehicle recovery profits
Also built
- Inventory Intelligence PlatformA weekly pipeline collapses a multi-million-row inventory history into one row per item and site, so an agent can answer any history question with a single small lookup.2025–2026
- Multi-Provider LLM Batch RuntimeA batch runtime across four model providers whose rate limits are discovered by a control loop rather than configured, and whose credentials never touch a developer's machine.2026
- Freight Mode-Shift Optimization EngineA weekly pipeline over planning, demand and lane-rate data that ranks mode-shift candidates by modelled freight savings and carbon impact.2025–2026
- Transit Time Prediction ServicePer-segment transit-time prediction by lane and shipping method, trained on eighteen months of closed orders and scored against the live backlog, with a validation gate that fails the pipeline if a join silently multiplied the rows.2026
What I do
Four things, and what each one actually means
Pick one. Each is a claim with the systems that have to back it.
Multi-agent systems that do not run away, do not leak, and do not lie.
Budgets at the tool-call boundary, wall-clock deadlines that survive a human-approval pause, durable checkpoints on evictable infrastructure, cross-instance run control, and authorization that fails static rather than open. Every one of those exists in my systems because something specific broke first.
- AI Compliance Investigation PlatformAn analyst names a party; six sanctions regimes, adverse media, aliases and corporate ownership come back as a cited, defensible risk determination.
- Enterprise NL-to-SQL Agent PlatformPlanners ask supply-chain questions in plain English; 32 domain agents answer with SQL-grounded results, under role-scoped access the model cannot talk its way past.
- Shipment Milestone Tracking AgentA shipment-tracking assistant rebuilt from a single-file prototype into the same platform architecture as the enterprise agent estate, including a middleware guard that strips SQL the prompt already forbade.
- Carrier Root-Cause Validation AgentValidates carrier on-time-delivery root-cause codes against a declarative decision tree, with deterministic routing before the model and a human ground-truth workbook behind the evaluation.
The through-line
These are not unrelated projects
Seven ideas recur across every system on this site. The number is how many of them instantiate it.
- 01Cheap deterministic checks before expensive judgement4
- 02Fail closed on ambiguity, fail open on infrastructure3
- 03Move domain knowledge out of prompts into reviewable artefacts3
- 04Guard at the boundary, not in the prompt3
- 05Measure the thing you are about to trust4
- 06Say what you cannot say3
- 07Pipeline first, then the agent on top of it4
How I work
Judgement, written down where it can be checked
Technical direction, cross-functional influence, and the decisions I would defend, including the ones that made my own earlier work look worse.