DutyBench
In progressAn empirical study of agent harness design: when does a model checking its own work add anything on top of deterministic action controls?
- Python
- pydantic
- Docker
- GitHub Actions
Software engineer building AI systems.
Most of what I build starts by talking to whoever's going to use it. The rest is the less glamorous half of the job: the data, the retrieval, and working out whether the thing actually holds up. Most recently that was an internal assistant at Spotify that turned most of an hour of digging into a short conversation.
I'd rather be on a small team where engineering, product and the customer are the same conversation.

An empirical study of agent harness design: when does a model checking its own work add anything on top of deterministic action controls?
A daily quest log that sets my priorities while there's no manager to set them: routine quests, plus a main quest with minimum, target and stretch tiers.
Retrieval-augmented QA over the FCA Financial Crime Guide, built to be measured rather than demoed. Master's dissertation.
85% retrieval hit rate
Early warning for sepsis in intensive care, built so the doctors using it can see why it flagged someone.
AUC 0.90 · recall 0.82
Conway's Game of Life rebuilt as a concurrent Go system, then distributed across AWS worker nodes behind a broker.
6.7x faster than single-threaded
Found the bottleneck by interviewing the engineers who would use it, scoped an MVP from what they said, then built the pipelines behind it and presented the architecture to 60+ engineers and stakeholders.
Python · BigQuery · Agentic AI
Built the docs portal for HSBC's internal component library, used by 200+ engineers, and added version history so teams still on older releases could find docs matching the code they were actually running.
React · TypeScript · Playwright
First Class Honours. Dissertation on retrieval architectures for financial regulatory question answering.
ML · Computer Vision · Distributed Systems