SME INBOX · SELF-HOSTED · MADE IN EUROPE
Every incoming email,
sorted or safely set aside.
mtlm-router turns a shared mailbox into typed decisions: devis, commande, litige, facture, spam-adjacent — each mail gets a route and a confidence, or an honest delegate when the right answer isn't clear. On-prem, on CPU, flat cost.
How it fits an integrator's stack
1 · Plug the mailbox
A small connector reads maildir / mbox / IMAP and posts each message to /v1/route. Output: sorted folders plus a decisions.jsonl audit trail — who was routed where, at what confidence.
2 · Train the client's head
Export a few hundred labeled mails from the client's archive, train a head in under an hour on CPU. The trunk never changes — the head file is the client's, portable, versionable.
3 · Let it abstain
Below the confidence threshold the mail lands in a review queue instead of the wrong folder. Wrong-folder mail costs more than unsorted mail — the engine is built to know the difference.
Why not just prompt an LLM?
Cost shape
An LLM call per mail is a meter running forever. This is a 7.8MB appliance on a VM — the hundredth mailbox costs the same as the first.
Calibrated no
A prompt improvises a label. This returns a typed decision with a confidence trained against a holdout — abstention is a first-class output, not a hope.
Data stays
Client mail never transits a third-party API. For an integrator selling to cautious SMEs, that's the whole sales conversation.
Honest scope
- ✓ Proven on a real production export: 79% holdout on 6,402 municipal requests — same engine, same recipe.
- ! Head quality is data-dependent. On a generic English email corpus (Enron), the same recipe scored 38–48%. Whether your mailbox transfers well is a measurable question, not a promise.
- ✓ So the pilot is the product: send a labeled export, we train the head and report the holdout number before you commit to anything.
Bring a mailbox export, get the number.
The path to a pilot is one labeled export away. We train the head and show you the holdout accuracy on your data first — then you decide whether to embed the appliance, under your brand, at your price.