On-device AI

Huma works for you

Real machine learning for phishing detection, reading, and prefetch — no model provider, no telemetry, nothing sent to a server.

Huma Guard

It catches phishing before you type

Guard is the layer that runs before every navigation and after every page loads. It works in two passes.

The URL model

Every navigation is scored by a real trained neural network — a small multilayer perceptron over interpretable features of the address (host entropy, look-alike characters, brand tokens in the wrong position, risky top-level domains, credential words, and more). It was trained by gradient descent on labelled examples, exported to ONNX, and runs in microseconds with tract. The exact same feature extractor is used for training and for scoring, so what the model learned is what it sees in your browser.

The content model

After a page loads, Huma reads its title and visible text. The decisive signal is brand impersonation with a domain mismatch: a page that dresses up as a brand you know, sits on a look-alike domain, and asks for your password is flagged — while that brand's real login on its own domain is not, because the domain owns the brand.

The warning

A flag drops the Dynamic Island from the top in amber, names the impersonated brand and the reasons, and gives you three choices: go back, report it, or allow it anyway. Nothing is blocked outright — Guard informs, you decide.

Self-adaptation

It quietly learns from you — privately

This is a small, honest analog of self-adapting models: your real browsing outcomes become tiny training signals the model applies to itself, on your device, forever offline.

You correct it

Allow anyway on a false alarm tells Huma that this shape of site is fine. Reporting a page, or a host later appearing on the threat feed, tells it the opposite.

It self-edits

Each outcome is one bounded gradient step on a personal adapter layered over the trained model — capped so it can only nudge, never take over.

It stays yours

The adapter lives in a local file and is never uploaded. Delete it and you're back to the shipped model. Nothing about your corrections leaves the machine.

Reading and prefetch

Huma Read

Summarize any long page on-device. The summary is extractive — built from real sentences on the page — so it's faithful and cannot hallucinate. Trigger it from the toolbar or the command palette; the result appears in the Dynamic Island.

Huma Predict

Vev learns your navigation patterns locally (a simple model over the sites you move between) and, when it's confident about where you'll go next, pre-warms that site's DNS, TCP, and TLS in the background so it opens noticeably faster. Weak or thin patterns trigger nothing, so there's no wasted prefetch.

Everything Huma does happens in the browser process on your device. There is no account, no model provider, and no telemetry.

FAQ

Questions about Huma

Does Huma send my browsing to a server?

No. Huma runs entirely in the browser process on your device. The phishing model is a ~1 KB file bundled in the binary and executed locally by tract (a pure-Rust ONNX runtime). There is no model provider and no network path out of the AI.

How accurate is the phishing model?

On a held-out test set it scores about 95.7% accuracy with 0.92 precision and 0.96 recall. It is a small trained multilayer perceptron over interpretable URL features, not a hand-written rule set, so it generalizes to new phishing that isn't on any blocklist yet.

Will it flag my real bank login?

No. The content check only flags a page when it presents itself as a brand AND the domain doesn't belong to that brand AND it's asking for credentials. A genuine login page on its real domain fails the first test, so it is never flagged.

What does 'learns from my choices' actually mean?

Each time you click Allow anyway on a warning, or a site later turns up on the confirmed threat feed, Huma applies one bounded gradient step to a small personal adapter stored in a local file. It nudges future scores for similar sites up or down. The adapter is capped so no single stream of feedback can run the model away from its trained baseline, and it never leaves your machine.

Is the summarizer generating text with an LLM?

No — Huma Read is extractive. It selects the most central real sentences from the page, so it is fast, private, and cannot hallucinate. Nothing is invented.