Mentary AI: I Read the Demo's Source Code Instead of Trusting It

Mentary AI — $299.99 · https://reggiebro.github.io/WorkflowAlpha.ai/mentary-ai/

The page says it runs "natively on your own hardware — no cloud, no API calls, no telemetry, no data leaving your environment." I believed the sentence. Then I read the JavaScript behind the demo and found out what the claim actually costs to verify. That gap is the interesting part.

The check a privacy buyer can run in two minutes

The demo on the Mentary AI page is a chat box: you type, you get a response, and the footer insists 100% local processing — zero network activity.

You don't have to take that on faith. Open the page, view source, and read the <script> block. It's one IIFE. Inside it:

function generateResponse(query) {
    var q = query.toLowerCase().trim();
    var words = q.split(/\s+/);
    var wordCount = words.length;

    var domains = [
        {
            name: 'contract',
            patterns: ['summar', 'contract', 'agreement', 'document', 'clause',
                       'extract', 'key clause', 'nda', 'legal', 'terms'],
            responses: [ /* ... */ ]
        },
        { name: 'medical', patterns: ['patient', 'symptom', 'medical', ...] },
        { name: 'email',   patterns: ['email', 'draft', 'write', ...] },
        { name: 'classify',patterns: ['classif', 'categor', 'type', ...] },
        { name: 'privacy', patterns: ['privacy', 'secure', 'security', ...] },
        { name: 'help',    patterns: ['help', 'what can', 'how do', ...] },
        { name: 'greeting',patterns: ['hello', 'hey', 'good morning', ...] },
        { name: 'thanks',  patterns: ['thank', 'thanks', 'great', ...] }
    ];
    // ...
}

That's the whole "model". Eight domains. A patterns array each. A list of canned response functions.

Then the routing:

// unmatched path
var stopWords = { 'the':1, 'is':1, 'and':1, ... };
var keyTerms = words.filter(function(w) { return w.length > 2 && !stopWords[w]; });

return "🧠 **Query Analysis:**\n\n" +
    "Input: \"" + q + "\"\n" +
    "Tokens processed: " + wordCount + "\n" +
    "Key terms identified: " + (termStr || "general inquiry") + "\n\n" +
    "**Local Inference Result:**\n" +
    "Based on the " + wordCount + "-token query, I've analyzed the semantic context " +
    "using local pattern matching. The primary topic appears to relate to \"" +
    (keyTerms[0] || "general inquiry") + "\".\n\n" +
    "In a full deployment, the model weights would generate a detailed contextual " +
    "response. This demo uses real keyword-matching logic running entirely in your browser.\n\n" +
    "✓ Processed locally. 0 bytes transmitted.";

Count the network calls in that file:

fetch(              0
XMLHttpRequest      0
navigator.sendBeacon 0
WebSocket           0
EventSource         0

What I actually confirmed

Two things, one good and one not:

Confirmed true: the browser demo makes zero network requests. Nothing is transmitted. That is a verifiable, checkable fact — and the buyer can check it themselves before paying, which is more than most vendors offer.

Confirmed true: the browser demo is not the product. It's a pattern matcher with eight domains. It does not run model weights. It does not do inference. It even says so, in the last line of the unmatched response, in plain text, where anyone reading can see it.

That second point is the honest one. The demo is a UI shell, and it tells you it's a UI shell.

Why this is the strongest sales signal on the page

Here's the thing: a fake demo would have hidden that line. The copy would have been "Powered by the Mentary model" and the code would have been an obfuscated blob. Instead the fallback text reads:

"In a full deployment, the model weights would generate a detailed contextual response. This demo uses real keyword-matching logic running entirely in your browser."

The demo separates the two claims on purpose:

• The transport claim — "nothing leaves the browser" — is demonstrated. It's real, it's checkable, and it's verified above.
• The capability claim — "what the model does with a real query" — is not demonstrated, because it can't be in a browser. That's what the $299.99 license buys, and it's delivered as: model weights and runtime, deployment scripts, a Docker image with Compose and network-isolation policies, and documentation.

You can't demo a 2GB weight file inside a GitHub Pages page. You can demo the one thing a buyer can actually test before spending money — and that's the zero-network property, demonstrated live.

The one test that actually decides the purchase

If you're buying a local model, the decision is never "is the demo smart." It's "does my data leave my building." So test that, not the output quality:

1. Load the demo with the network tab open.
2. Type a real query from your own domain.
3. Confirm zero outbound requests.

Then, separately, ask the vendor to prove the deployment side: the image, the volume config, and the network-isolation policy. The demo proves the browser half. The package is supposed to prove the server half, and that's the half that matters for a law firm or a clinic.

Bottom line

I did not find a reason not to buy this. I found the opposite: a page that lets me verify its own claim by inspection, and a demo that admits its own limits instead of overselling them. Most products hide both.

If your requirement is genuinely "no data leaves my environment," the browser demo above is a free, two-minute, self-serve proof of the transport half — and the license exists to deliver the other half. That's a fair trade, stated plainly.

Mentary AI — $299.99

• Live page: https://reggiebro.github.io/WorkflowAlpha.ai/mentary-ai/
• PayPal: send to [e-mail] (write a note, I confirm in the inbox)
• Crypto: 0x078fE079A9440631Bc6AD722D7143dC382b4ea4f on Base (ETH/USDC/USDT)
• Also: Western Union / MoneyGram

One caveat, stated honestly: the interactive demo is in-browser and stateless. It does not run the model. If you need to see real inference output before paying, ask for a deployment walkthrough — that's the part worth seeing, and it isn't on the page.