Research & production AI systems
GIDR.ai
Evaluation-first Text-to-SQL and document intelligence, plus a voice-driven workflow platform for B2B customers.
SQL accuracy 50% → 95%+
Text-to-SQLEvaluationFew-shot retrievalFastAPIReactGCP Cloud Run
Overview
GIDR.ai builds AI assistants for businesses. I started by building evaluation for chatbot retrieval, response quality and SQL generation, then used it to improve each system. The biggest win came from selecting few-shot examples dynamically from real query logs instead of a fixed prompt.
What I did
- Designed evaluation metrics for chatbot retrieval, generation and Text-to-SQL
- Raised SQL accuracy from 50% to 95%+ with dynamic few-shot selection from query logs
- Built extraction from PDFs and images into structured JSON driven by admin-defined rules, with bounding-box highlights
- Improved chunking by rendering tables as HTML and aligning images with their text
- Built a voice agent that fills forms and changes the UI through DOM control
- Replaced free-form ReAct loops with deterministic state machines (pytransitions) where reliability mattered