LocalFlow – AI data analysis where the LLM never sees your data

Name: LocalFlow – AI data analysis where the LLM never sees your data
Availability: InStock
Author: renaudpawlak

by renaudpawlak·Jun 2, 2026·2 points·0 comments

Visit Project View on HN

AI Analysis

●●SolidBig BrainSolve My Problem

Sends schema not data to LLM, unlike Vanna, but browser performance millions rows remains dubious.

Strengths

•Explicit metadata boundary prevents raw data leakage during inference calls completely.
•Generated code executes deterministically locally, enabling zero-token cost scaling for queries.
•Clean npm package with live demo for immediate CSV testing today.

Weaknesses

•JavaScript execution on millions of rows in browser will likely bottleneck.
•Competes with established Text-to-SQL tools like Vanna.ai and Defog in crowded market.

Post Description

Hi HN, I'm the author of the JSweet project (Java-to-TypeScript transpiler), but today I'd like to show you something new: LocalFlow.

There are two common ways to analyse your data with AI ("Talk to your Data"): 1. Use powerful frontier models on the cloud — but hand over your data and pay per token 2. Use on-premise or local models — but accept weaker results and significant hardware costs

LocalFlow proposes a third approach I call "metadata-first AI": send only the schema to the LLM (column names, types, a few stats — never the actual values), and ask it to generate code that runs locally on your full dataset (scale it to millions of rows at zero additional token cost).

I'm curious what you think :)

Live demo (drop in any CSV): https://apps.localflow.fr/demo/ GitHub: https://github.com/localflow-ai/localflow-core