Synapse

Inspiration

Text-to-SQL can be used to answer ad-hoc data analysis questions and solve self-serve analytics

What it does

Synapse connects to tables in Google BigQuery, converts natural-language queries into SQL queries, and returns answers from the database. It summarizes learnings from the conversation based on feedback given by the user and through self-correcting errors it encounters.

How we built it

We connected analytics data in Google Bigquery with a Google Vertex AI-powered chat app in Firebase and LLM monitoring using Braintrust, along with the option to use speech-to-text through Modulate

Challenges we ran into

Connecting the tools together.

Accomplishments that we're proud of

Getting intelligent learnings from the model and automatic LLM-as-a-judge scorers through Braintree

What we learned

Metadata and feedback are immensly helpful for improving LLM apps

What's next for Synapse

Refining metadata available to the agent about the analytics datasets to generate more intelligent analytics and learnings. Also, embedding BI tooling into the chat app.