Tines The single, secure environment for agents, apps, and automations.
Promoted
Maker
📌
Hey PH! 👋 Tsing here, maker of LakeMind.
I've spent the last couple of years building and using "chat-with-data" tools, and kept hitting the same wall: every one of them wants you to upload your warehouse to the cloud, talk to a server-side model, and pray the first SQL it writes is correct. For anyone with real production data, that's a non-starter.
So I went the other way. LakeMind runs entirely on your machine. Your files and databases never leave your laptop — a local DuckDB engine acts as the coordinator, heavy aggregations get pushed down to the source database, and only schema plus a handful of sample rows ever reach the LLM. No data warehouse, no ETL pipeline, no upload.
The second bet is on how AI should work with data. Most tools treat the LLM as a one-shot compiler obsessed with writing the perfect SQL on try #1. We treat it as an explorer: a local millisecond feedback loop lets the agent try → hit an error → self-correct, many times per second, against real data. In practice this means cheap fast models (think deepseek-v4-flash) can do surprisingly serious work, because the loop — not the model size — does the heavy lifting.
A few things it can do that I'm proud of:
Join an Excel file with a Postgres table without moving either side anywhere.
Materialize agent-produced result tables (t_ / v_ / tmp_) back into the local lake, so analysis actually accumulates instead of vanishing.
OKF — an open Markdown/YAML format that travels with the data, so when you hand a dataset to a colleague, the semantic context (joins, metrics, column meaning) rides along. No cold start.
It's a desktop app built on Tauri + Rust, currently macOS-first (Apple Silicon), with Windows/Linux on the roadmap.
I'd love your honest take — especially on where the "local AI analyst" story still feels rough. What would make you trust an AI agent enough to point it at your real database? Happy to dig into architecture, the federated pushdown design, or anything else in the comments.
