AI Objectives Institute · 2023 - 2024

Talk to the City

I built Talk to the City, an LLM interface for collective decision-making, and onboarded Taiwan's Ministry of Digital Affairs.

Svelte
Firebase
Python
UMAP
HDBSCAN
BERTopic
A pipeline in Talk to the City Turbo processing a survey on same-sex marriage, in Traditional Chinese. 4:52.

What I did

  • Redesigned the React prototype as a Svelte and Firebase app, with LLMs processing topics and subtopics in real time.
  • Built a dependency graph of 50 processing nodes, so each dataset gets its own pipeline topology.
  • Took over the Python backend for argument and topic extraction: UMAP, HDBSCAN and BERTopic.
  • Wrote 200+ automated tests for the app and its pipelines.
  • Built synthetic datasets as reference pipelines, then refined them against real data.

Links