Research
The engineering decisions behind Warren: voice, memory, financial modelling, UK-specific knowledge, product guardrails and background monitoring.
Most financial planning apps go silent after onboarding. We built an AI agent that reads each user's full financial context every day - and stays quiet unless it has something genuinely useful to say.
Dima Tarasenko
9 min read
Our first financial model was a unique piece of LLM-generated Python for every user. It worked - until it didn't. We analysed 1,682 production models, found common patterns and built a deterministic engine to replace them.
Off-the-shelf LLMs are not designed with UK financial regulations in mind. We gave Warren a source of truth for UK financial rules - 746 concepts, 7,753 atomic units, all sourced from gov.uk.
