deepabi.md

DEEPABI

Full-stack development
A specialized AI assistant that makes working with databases feel like a conversation. Ask a question in plain language, and DEEPABI reads the schema, plans the query, runs it, and hands back the answer.

The problem

Most people can't get answers out of their own data. SQL is a wall, and the person who can write it is always busy, so questions pile up and decisions wait on an engineer. DEEPABI takes that wall down. You ask in plain English, and it does the translating.

What it does

Capability
What it means
Natural-language querying
Ask in plain language instead of writing SQL
Schema-aware planning
Understands your tables and relationships before it answers
Multi-database support
Works across relational and other database systems
End to end
The chat interface, the API, and the engines behind them

How it works

A question comes in as plain text. DEEPABI grounds it in the live schema (table names, columns, relationships), uses an LLM to plan and generate the query, runs it against the database, and returns the result in a readable shape. The model does the understanding and generation, and guardrails around execution make sure a question can't turn into damage.

Highlights

  • Natural-language interface over relational and other database systems
  • LLM-powered query understanding and generation
  • Built end to end: interface, API, and the engines behind them

Stack

TypeScript TypeScript · Next.js Next.js · Node.js Node · PostgreSQL PostgreSQL · Anthropic OpenAI / Anthropic

Why this stack

Choice
Why
Instead of
Next.js
One framework for the chat UI and the API routes behind it
a separate frontend and backend to keep in sync
Node
A natural fit for orchestrating LLM calls and streaming responses
a heavier runtime for what is mostly I/O
PostgreSQL
A strong, well-understood target to reason about and query
starting with a store that lacks a rich query language
OpenAI / Anthropic
Best-in-class language understanding for turning English into intent
training a bespoke model for a general problem

The AI part was the easy bit. The real work was trust. That's the whole idea here: let your data answer you, without ever letting it lie.