Making data accessible using LLMs

On my visit to TU Berlin to talk about data accessibility and LLM-powered systems.

Back to Berlin

This winter, I was invited to present my research at TU Berlin and BIFOLD. Guest talks are a nice forcing function: you need to step back from coding and explain why the problem matters in the first place. In my case, the goal is quite simple to state: people should be able to query their data without first becoming experts in data systems.

Data systems beyond chatbots

Large language models make it tempting to imagine that every data problem can be solved by adding a chat interface. Natural language is certainly useful, but accessibility requires more than translating a question into code. A system must also find the relevant data, combine heterogeneous sources, execute a reliable pipeline, and make sensible trade-offs between quality, runtime, and cost.

My talk, Making Data Accessible using LLMs, focused on this broader systems perspective. I discussed how declarative interfaces can make complex AI pipelines easier to express, and how data management techniques can help optimize the expensive and non-deterministic operations behind them.

BIFOLD shared a short summary of the talk. Many thanks to Ziawasch Abedjan, Sebastian Schelter, Matthias Boehm, Volker Markl, and the whole group for the invitation and the interesting discussion after the talk!