AI agent for retail analytics
“Achieved 90%+ accuracy on SQL generation running locally on a laptop CPU with a 3.8B parameter open-source model, handling hybrid RAG+SQL queries for retail analytics without APIs or paid services”
solo-developer
Why they built it
To test how far small open-source models can go when treated like a complete system rather than simple prompt-response tools
What worked
High performance of small open-source models when architected as a full system (90%+ SQL accuracy on CPU), hybrid reasoning combining RAG and SQL, self-correction capabilities, full traces for transparency
What broke or was painful
Not explicitly detailed; initial challenges with model context awareness implied before optimizations like proxy datasets
The result
Achieved 90%+ accuracy on SQL generation running locally on a laptop CPU with a 3.8B parameter open-source model, handling hybrid RAG+SQL queries for retail analytics without APIs or paid services