SyntheticMind — Private on-device AI

An AI that's yours alone

Think about it: everything you tell a regular AI assistant travels to someone else's server. Your doubts, your contracts, your worries. SyntheticMind was born from refusing that: a full AI that lives inside your iPhone and answers even with no internet. Turn on airplane mode and ask it something — that's the proof.

A legal question?Ask about Spanish employment law, inheritance or any article: it answers with the actual text of the official gazette (BOE) and shows you where it comes from.

No accounts, no passwordsOpen it and it works. Nobody even knows you use it.

It shows its cardsEvery answer carries a seal with the sources used and their confidence. And when it doesn't know, it says so.

Remembers what mattersIts memory lives on your phone, not in a cloud. Erase it whenever you want.

Snap a photo and askA note, an invoice, a contract: it reads and understands them without leaving the device.

Talk to itDictate your question and listen to the answer.

🔧 For the curious: how it works insideArchitecture, engine and technical decisions — only if you feel like diving in▾

Two platforms, one engine: the Rust cognitive core compiles natively for iOS and to WebAssembly for the browser, with the same verification guarantees on both.

Rust cognitive core. Multi-layer orchestrator with reasoning, memory, understanding, planning and safety modules. Bridged to Swift via UniFFI.

On-device inference. A fine-tuned Qwen model running on llama.cpp with Metal acceleration on iPhone; Qwen 2.5 via WebLLM + WebGPU in the browser.

Local knowledge base. SQLite in WAL mode with encryption at rest: 5,300 articles of Spanish law parsed article by article plus 9,000 regional open-data records, ingested with idempotent pipelines.

Anti-hallucination grounding. Explicit-citation detection ("article 348 bis" → official text), term-overlap retrieval with thresholds, and a grounding-strength report that decides between a grounded or a cautious answer.

Output verification. A verifier checks the answer's citations against the real corpus and flags non-existent ones before anything is shown to the user.

Zero-cost web version. The same grounding engine compiled to WASM with wasm-bindgen, the corpus as static JSON, an installable PWA and GitHub Pages as the only "server".

LanguagesRust + Swift (+ WASM)

ModelFine-tuned Qwen, on-device

Corpus14,000+ official documents

Data collectedNone

Cloud cost$0

Web infrastructureGitHub Pages (static)

On screen

Get it on the App Store Try it in your browser