I turn fuzzy user problems into AI products people use: I define the product, write the spec and the evaluation criteria, and ship it with an AI-native workflow.

MSc Management @ Imperial College London · BSc Statistics, Economics and Finance @ UCL · London

jev-chat: global edition · Co-creator, Product Lead (Global)
An open-source chat decision assistant . Jev reads the chat, works out what the other person wants, and drafts replies for you to choose from. You press send. I rebuilt it from the ground up for English and overseas messaging and shipped Jev for WhatsApp on Android. Telegram and other major platforms are next.

Flowtype · Creator
Local-first voice dictation for macOS: on-device speech recognition, paste on release, and a math mode that dictates LaTeX. I built it because cloud dictation tools were slow and garbled mixed Chinese–English speech.

codex-chat-keeper
A local macOS app and CLI to back up, repair and migrate Codex Desktop chat history.

• Start from the user's problem, not the model. Write the product principle down before the prompt.
• Make AI behaviour testable: fixed judgement questions, explicit scoring rules, and tests that lock them in.
• De-risk the biggest unknown first, then ship small and test on real devices.

LinkedIn