StructAI

The story every analyst knows: someone hands you a spreadsheet, the columns are named Cust_Email_Final_v2, the same customer appears with three slightly different addresses, and you spend an afternoon writing one-off Python before you can even start the work you were actually asked to do.

StructAI is the tool I wanted that afternoon. You drop a document on a project, an agent profiles it, proposes a Postgres schema (extracted entities, foreign keys, primary keys, the boring choices made for you), and once you accept the schema it generates and runs the import script inside a single transaction. Every run takes a database snapshot first so you can undo any import that landed badly.

What it does well

• Schema-first review. The agent shows you the DDL before it writes a line of import code. You accept, or you reply in natural language (“split the address into its own table; make customer.email the PK”) and it revises. This is the most expensive decision in an import, and it’s the one the model should not silently make.
• Self-correcting execute loop. When the generated script blows up (encoding, mixed date formats, an embedded comma in a quoted field), the agent reads the stderr tail, diagnoses the root cause, and rewrites. Five attempts max; past that, it asks for help instead of thrashing.
• Cheap undo. Every successful import gets a template-DB snapshot. One click rolls the project back to before the import landed. Snapshots are pruned by a retention sweeper.

What it’s made of

• FastAPI + arq on the backend: one job at a time, server-sent events to the UI so the agent’s thinking shows up in real time.
• React + Vite on the frontend, with an interactive ER diagram and a data browser that does server-side sort and per-column filters.
• Postgres as both the metadata store and the per-project data DB, where template-DB clones make snapshots near-free.
• Anthropic Claude under the hood, called through a small agent-loop wrapper that lets the model ask the user mid-stream when it has to make a judgment call.

What I learned

• The right place for the human is before the agent has done expensive work, not after. Reviewing a schema takes 30 seconds; reverting a bad import takes ten minutes of cleanup.
• Letting the agent ask for clarification is much better than telling it to guess “sensibly”. The model will happily invent a heuristic; what you wanted was for it to say “I don’t know, you tell me.”
• A small bounded retry loop (“try, read the error, fix, try again, give up at 5”) catches almost every real failure. Unbounded retries are how you get $40 of tokens spent on a CSV with a misspelled column.

Try it at structai.adityaviki.com.