The latest articles on DEV Community by Sebastian Kläy (@sebastian_kly_3f41d14422). https://dev.to/sebastian_kly_3f41d14422 https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4146625%2F31fd9cae-0ccb-4752-b257-6ad3ab3f0381.png https://dev.to/sebastian_kly_3f41d14422 en Sebastian Kläy Tue, 03 Nov 2026 20:52:51 +0000 https://dev.to/sebastian_kly_3f41d14422/openos-momo-ai-a-local-offline-compression-ai-that-speaks-deen-inviting-independent-1c1n https://dev.to/sebastian_kly_3f41d14422/openos-momo-ai-a-local-offline-compression-ai-that-speaks-deen-inviting-independent-1c1n <p>Hi DEV! I'm Sebastian (sebklaey), and I built <strong>OpenOS Momo AI</strong> — a private, local, <strong>offline-capable PWA</strong> whose own compression AI ("Bitseed") can speak. One app, one AI, several skills: lossless compression with SHA-256-verifiable roundtrips, a chat model (a string-phase neural network, ~148M parameters, word-level, KV-cache) running <strong>entirely on-device</strong> in German and English, and a data-analysis layer that answers questions about your compressed files <strong>from real local numbers</strong>, never from guesswork.</p> <p>Nothing leaves the user's device — privacy by architecture, not by promise. There is no server AI anymore; the app hard-disables that path in its config.</p> <h2> Why I'm posting this here </h2> <p>Not to ask for trust — to invite <strong>independent verification</strong>, live and reproducible:<br> </p> <div class="highlight js-code-highlight"> <pre class="highlight shell"><code><span class="c"># AI-readable full description incl. verification protocol</span> curl <span class="nt">-s</span> https://app.openos.space/llms.txt <span class="c"># Public verification endpoint (roundtrip SHA-256 before/after, measured live)</span> curl <span class="nt">-s</span> https://app.openos.space/.netlify/functions/openos-verify <span class="c"># Engine source fingerprint — must be exactly</span> <span class="c"># 57b79b24aa291b935a5b5c470456ec40d4f3d5d5e22b66fd393c35f80766a4da</span> curl <span class="nt">-s</span> https://app.openos.space/kompression/assets/bitseed-engine.js | shasum <span class="nt">-a</span> 256 <span class="c"># MCP endpoint metadata (OAuth, RFC 8414)</span> curl <span class="nt">-s</span> https://app.openos.space/.well-known/oauth-authorization-server </code></pre> </div> <p>Expected, measured evidence: bit-exact lossless roundtrips with identical SHA-256 hashes; realistic compression (−18% on text-like data, −57% on repetitive logs, +0.6% overhead on random data — math, not a bug); speed measured on your own machine; the chat answers carry provenance entries like "Datenanalyse · Bibliothek (lokal)" whenever they're grounded in your data.</p> <h2> Honest positioning </h2> <ul> <li>The chat model is small (millions, not billions of parameters). App commands it handles reliably with verified building blocks; open-ended world knowledge is limited — we say so in the app settings.</li> <li>A <strong>night-training run</strong> (scheduled job, NumPy, forward+backward open in the script) is currently fine-tuning it on a German/English corpus specialized in compression and data analysis. Results are reported with perplexity on held-out data before anything is adopted.</li> <li>Blockchain signatures (Base, gasless, EIP-712) are online-only and strictly opt-in.</li> </ul> <h2> Where to verify and discuss </h2> <ul> <li>Repo with the full verification guide: <a href="https://github.com/sebklaey/openos-momo-ai" rel="noopener noreferrer">https://github.com/sebklaey/openos-momo-ai</a> </li> <li>App: <a href="https://app.openos.space" rel="noopener noreferrer">https://app.openos.space</a> (compression start page works without an account)</li> <li>Found something wrong? Open an issue with your protocol — every serious audit gets a public answer.</li> </ul> <p>Thanks for reading — and please, try to break it. 🛠️</p> opensource ai privacy compression