🦅 ReverseML Mini — how it works
A private, self-adapting on-device AI. It learns you, tunes itself to your device, and works with your documents — offline.
Every number below is measured and reproducible — not a mockup.
1 · It learns you
Same model, same questions — with memory OFF it answers generically; as its private on-device memory fills over a week, personalization rises to 92%. Mechanism: private memory + retrieval — no weight fine-tuning. The model never leaves your device.
Memory OFF: 0% → Day 1: 100% → Day 5: 90% → Day 7: 92% · memory recall 100%. Same model. Same questions. Personal memory makes it yours.
2 · It self-tunes to your device
One engine. On startup it reads the device — CPU / RAM / battery / thermal / accelerator — and picks a run mode with real runtime settings. Same AI engine, different device, automatic optimization.
| Device | Chosen mode | Context | Threads |
|---|---|---|---|
| Low-end phone (3 GB) | Tiny | 1024 | 4 |
| Typical phone / laptop | Balanced | 2048 | n-1 |
| Desktop PC (32 GB) | Long-memory / Turbo | 4096–8192 | all |
| Snapdragon (NPU) roadmap | NPU Balanced | 2048 | n-1 |
| Low battery | Battery Saver | 1024 | n/2 |
| Overheating | Sustained (thermal) | 1024 ↓ | throttled |
Live adaptation: NPU Balanced → (84 °C) Sustained → (12% battery) Battery Saver — automatically, all offline.
Real, not printed: same model + prompt, 1→4 threads measured 16 → 29 tok/s (~1.8×) — the knobs change actual execution. NPU / larger-model tiers are marked roadmap and are never shown as running.
3 · It works with your documents
Attach a PDF / Excel / Word / DXF, ask a question — the answer comes back with its source (page / sheet-cell / paragraph / CAD layer), extracted on-device.
- PDF manual → “How often must the seal be replaced?” → Every 8000 hours · source p.4
- Excel sheet → “Flow rate at t=30 s?” → 450 L/min · source Measurements!r5
- Grounded & offline: PDF 7/7 · Excel 6/6 answers correct with correct source in tests.
Not a chatbot. A self-adapting private on-device AI engine.
It learns you, tunes itself to your hardware, and keeps working offline.
Get ReverseML Mini →Honest scope: it’s a ~1.5B model — private personalization, grounded documents and device self-tuning, not frontier-level reasoning. Embedded/autonomous use is an advisory layer with a human in command, not a certified autopilot.