ProjectBaseball.orgBuilding the First Dataset
Creating a list of natural language queries about Retrosheet data to analyze baseball data, mapped to SQL by an LLM.
Technical Blog
Designs, Architecture, and Code. LLMs, Evals, and Observability.
We’re trying something different. Instead of writing code, we write a detailed system plan. Then we use AI to write most of the code. We integrate the AI-generated code - it isn’t about AI building entire apps on its own. We write requirements and systems architecture; AI generates code; we integrate and test. Inspired by concepts in Andrej Karpathy’s Software 2.0 post
Offline Architecture v0
Training workflow for the RAG and prompting system.
System Architecture v0
Covers the e2e training, evaluation, logging, and user facing flow.