What Can LLMs Do with Messy Data?
summary
Evaluates what large language models actually do with ResStock-style data using pandas/DuckDB workflows and representative ResStock runs (solar thermal in Texas, HVAC in the Southeastern U.S.). Shows sampling, reshaping, and a schema “data card,” plus a deployed multi-agent pipeline—multi-agent orchestrator, statistics, visualization, and query agents—backed by a SQLite back-end and local open-source LLMs. Delivers generated/refined queries, plain-English impact explanations, and reproducible failure modes: token cost, hallucinated columns, and remaining need for human science judgment.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 1:46AI in energy research
- 3:30Simulating policies with Restock
- 7:38Messy energy data challenges
- 9:25From scripts to AI analysis
- 11:11GPT for data exploration
- 14:00Collaborating with Red Hat
- 17:09Multi-agent conversational AI
- 20:39Results and improvements
- 24:11Live demo
- 26:16Tool goals and limits
- 28:52Conclusion and Q&A
related talks
22mIntegrating LLMs with Traditional Data Analysis
33mDesigning LLM Applications for Massive Datasets
38mLLMs for Self-serve Analytics - Lessons Learned
41mBuilding LLM Agents – Challenges and Learnings
15mLLMs for Data-Driven Decision Making in Product Development
24mDemocratizing data access through LLM
49mA Practical Approach to Using LLMs: Peculiarities and Pitfalls
27mCan LLMs do research?