Immigration Assistant

We built the proof of concept as a small team, deliberately mixed: another data scientist and me on the system, a designer on the conversation and the interface, and two project managers keeping a tight timeline honest. The pace was fast from the first week, and it stayed that way. The interface was written in TypeScript and React, the retrieval and orchestration with LangChain over a LanceDB vector store, and everything ran on AWS Lambda, so there was no server to mind and the running costs stayed close to nothing. The testing was the team's finest hour, and I say that as the one who watched it from the engineering side: my colleagues took the assistant onto the streets and asked people to use it there, with their own questions, thinking aloud while they listened. That feedback shaped the answers, the tone and the interface directly. The response was warm, and the work drew real attention from government bodies and aid organisations, who saw what it could do for the people they serve. After the proof of concept I was pulled onto another client project, the Global Response Documents. The team carried the assistant forward, built it out and founded a start-up around it. Some time later they stopped: once they ran the numbers on what a large language model costs per question at the volume such a service would attract, together with the complexity of the material, the economics did not hold. I think that is worth saying on a portfolio. An AI product lives or dies on the cost of each answer, so it pays to do that maths early, ideally before the first line of code, and again as soon as real people are using it. And the maths keeps changing: prices per token keep falling, and capable open models that run on hardware of your own keep arriving. An idea that does not add up today may add up soon, and the work of getting the answers right does not go to waste.