JevMade

Sign in
← Back to experiments

Benchmarks & research

typesafe-local

typesafe-local asks a local MLX model typed questions and returns calibrated probabilities without generating or parsing prose.

Source screenshot of typesafe-local
SOURCE SCREENSHOTFull screenshot ↗

What it does

Built for Apple Silicon, it reports roughly 200 milliseconds for four questions over one document on an M4 Pro.

How you can use it

Start by collecting real customer tickets with known answers. For each ticket, list the yes-or-no checks, urgency ratings, and departments to test. Keep each choice list under 27 items. Always keep the list options in the same order so the results remain consistent.

A developer can run this code locally on an Apple computer. It works entirely offline without an account. Your app sends a ticket and asks several questions at once. Have your developer check the reported certainty numbers so unclear tickets can wait for a human review.

Maker-reported (not independently measured by JevMade): Four questions over one document: ~200 ms on an M4 Pro · Qwen3-1.7B, full validation split, bootstrap 95% intervals:

Primitives
choice, score, noul
Platform
Python
Added
Project created
GitHub stars
5 (snapshot, not live)

Keep this for later

Sign in to bookmark experiments, guides and videos, and keep notes only you can see.

Continue to sign in

We’ll bring you back to this listing.