millfolio
A private vault for your documents. Index your PDFs, Word docs, CSVs, and notes and ask open-ended questions — answered locally, your data never leaving the machine.
Privacy is protected by design, not guaranteed: only an aliased, de-identified schema reaches the frontier model — never your raw data — but no software is perfectly secure. See the disclaimer.
Isolating your data while running on the modest compute available on most people's machines is an ongoing, interesting challenge. The solution rests on three moving parts:
- The frontier model writes the code that runs on your data; it never sees the data itself.
- The local model processes your data as instructed by the generated code.
- A flexible registry of derived tags lets the vault scale past a few hundred records — computed from deterministic rules or, for fuzzier cases, ML prompts applied incrementally in the background via a materialization ledger (see the deep dive).
You'll get the most out of it if you review and edit your tags occasionally — the vault ships with sane defaults, but you know your data best.
This application has been tested on M2 and M4 hardware with 16–24GB memory.
Documentation →
Populate your vault, then organize it with keyword tags, AI tags, and tag groups.
The inference server →
The from-scratch, pure-Mojo GPU engine — socket to logits, every kernel custom-written — the models it serves, and the performance numbers.
The tools →
The from-scratch Mojo libraries underneath — networking, JSON, PDF extraction, vector search.
The Enclave →
How a frontier model can help with your data without ever seeing it.
A question becomes a program →
The deep dive on the codegen flow: a frontier model writes code that runs on your data, never the data itself.
Measuring quality →
How we quantify what int4 quantization costs, with perplexity.