Inspiration
Data is the backbone of every modern business, but it's locked behind a language most people never learn. We built SQL Whisper to break down that barrier — so anyone can ask questions of their data and get answers that drive real decisions. We wanted to provide a low-latency, voice-driven interface that allows users to translate natural language into actionable database insights without compromising the security of their data.
What it does
SQL Whisper leverages AI to turn plain English prompts into SQL queries and easy-to-understand responses giving insight of relevant data to the user. It keeps database authentication secure, by passing a unique ID to the AI Agent, which corresponds to credentials stored in our secure database. This way, credentials are never viewed by the frontend or the agent. The tool also provides visual learners with tools to browse their database. Our AI Agent is able to generate graphics for visualizable data including pie charts, line charts, and bar charts, all from a simple text prompt.
How we built it
We built SQL Whisper around a core principle: the AI agent should never touch raw data directly. When a user asks a question, the agent first reads a compressed snapshot of the database schema, writes an SQL query, and calls a secure backend tool to execute it. This keeps the context window lean regardless of database size, and means the agent reasons about structure rather than drowning in rows. On the security side, credentials are encrypted at rest and never exposed to the frontend or the agent, so every request is resolved server-side using an opaque connection ID. The result is a system where Claude iteratively plans, queries, and refines its way to a plain-language answer, without ever having direct access to the underlying data.
Challenges we ran into
- Due to time constraints, we stored our database locally on a member's computer, which made agentic workflow testing from other machines difficult. Our project also had a small number of files, so communication to prevent merge conflicts was key.
- We took care to ensure a smooth user experience. This included developing challenging and complex animations to ensure smooth transitions between different page elements.
- AI models have a tendency to hallucinate information, especially when dealing with large datasets. We had to include special safeguards to prevent this such as preventing the agent from attempting to create a line chart when presented with non-numerical data, for instance.
Accomplishments that we're proud of
We were able to build a smooth, clean implementation of several files operating on different programming languages in a short amount of time. We implemented a relevant process agentically by using Claude to make SQL queries.
What we learned
We learned about SQL querying and agentic workflows, as well as building API endpoints that are accessible and easy to use for frontend developers.
What's next for SQL Whisper
Our immediate roadmap includes migrating our local testing environment to a fully cloud-native architecture. That way, we can ensure consistent uptimes and accessibility for users.
Log in or sign up for Devpost to join the conversation.