Machine Learning Engineer Intern at Department of Statistics, May 2023 – Aug 2023
Prototyping LLM tools for public-sector analysts
LLM prototypes for government departments, covering text classification, Q&A and report generation, built with LangChain and GPT-3.5 in 2023.
The project
In mid-2023 I interned at the Department of Statistics, building LLM prototypes with different departments. GPT-3.5 was new, LangChain had just come out and people were already calling it bloated. We tried three kinds of tool: text classification, question answering and report generation.
What I built
One was a chatbot for the questions SingStat gets again and again, such as what a statistic means or where the data comes from. Another helped analysts who prepare quarterly sector outlooks for the Economic Development Board, and who had to pull together PDFs, news articles and CSV files to do it.
Both used retrieval-augmented generation: find the relevant documents first, then have the model answer from them.
Looking back
Most of this is standard now. LangChain, the library everyone called bloated that summer, runs a lot of production systems today, including some of the agents I work on at Carro.