Founding AI Product Engineer, NSAI
NSAI builds trustworthy domain-specific AI for buyers in regulated domains. Neuro-symbolic by design: the neural side proposes, the symbolic side disposes, and nothing ships without provenance back to the source document.
What NSAI is building
NSAI builds domain-specific AI for organisations in regulated domains, the ones that cannot act on an answer they are unable to trace back to a source. The design is neuro-symbolic: a language model proposes, a rules layer disposes, and nothing reaches a user without a citation back to the document it came from.
I work across both halves. I write the documents that decide what gets built, and then I build a large part of it.
What I have shipped
An LLM information-extraction service, built and launched from scratch. It replaced a rule-based engine and raised usable extractions by roughly an order of magnitude on identical input.
The evaluation and monitoring layer. The model has to answer in verbatim quotes. The service then independently re-locates each quote in the source document, so accuracy is measured per run rather than assumed, and a human review step handles adjudication.
An end to end LLM fine-tuning pipeline, FastAPI and Unsloth on GPU infrastructure, from job orchestration through QLoRA training to a quantised GGUF export for local inference. Mine from design through implementation and iteration.
A model selection study. Benchmarked base models from 3B to 24B parameters on accuracy against cost per run, which turned model choice into an explicit cost and performance decision rather than a default.
The services and interfaces underneath all of it: containerised FastAPI services, reusable Python modules, and React and TypeScript interfaces for the platform's data and model workflows.
What is not on this page
This is a private programme. The customer, the material it runs on, the commercial figures and the internal design are not mine to publish, so they are not here. Ask me in an interview and I will go as deep as I am permitted to.