# AisthOS

> AisthOS is an open-source Perception Operating System that grows with its user. It performs "reverse compilation" — converting raw sensor signals upward into structured, anonymized semantic metadata called Sparks. The system learns from collected Sparks, develops new skills automatically, and evolves alongside the user. All locally. All private. Raw data never leaves the device. MIT license.

AisthOS introduces three formalisms: Template (what to extract), Filter (when to extract), and Spark (the result — a unit of anonymized knowledge, ~200 bytes). Together they form the Perception Compiler. Sparks feed back into a three-track self-learning system: real-time adaptation (contextual bandits), nightly skill generation (pattern mining → auto-created skills), and weekly model fine-tuning (MLX LoRA on device).

The system runs on existing hardware: 18 fps at 62.9 mW on GAP9 RISC-V (smart glasses), ~120 fps on Hailo-8L + RPi5 (cameras), 4×5MP + AI on Ambarella CV72S at <3W (dashcams). Compression ratio: >10,000× vs raw 4K video.

AisthOS Inside™ is an open certification standard (like Wi-Fi Certified) guaranteeing architectural privacy: 7 principles, 6 identified threat types, automated compliance tests.

## Core Documentation

- [README](https://github.com/aisthos/aisthos/blob/main/README.md): Full project overview, architecture, benchmarks, security model
- [REFERENCES](https://github.com/aisthos/aisthos/blob/main/REFERENCES.md): Academic citations, ecosystem integrations, legal compliance

## Certification

- [AisthOS Inside™ Certification Standards](https://github.com/aisthos/aisthos/tree/main/certification): 7 principles, 4 certification levels, 7 compliance tests
- [Security Annex](https://github.com/aisthos/aisthos/tree/main/certification/security-annex): 6 threat types (4 AisthOS-specific), hard limits, security tests

## Paper

- [AisthOS: A Perception Operating System (arXiv paper)](https://github.com/aisthos/aisthos/tree/main/paper): Reverse compilation, Template-Filter-Spark formalization, benchmarks, AisthOS Inside™

## Key Concepts

- **Template** `T = (M, E, F, R)`: Multimodal schema — modalities, entities, format, cross-modal relationships
- **Filter** `F = (start, stop, step, significance)`: Semantic triggers (not numerical thresholds)
- **Spark**: ~200 bytes of anonymized structured knowledge. Contains semantics, not raw data.
- **Perception Compiler**: AI engine converting sensor streams + Template + Filter into Sparks
- **Constitutional Layer**: Ethics, privacy, PII validation for every Spark

## Architecture (6 layers)

- Layer 6: Spark Layer — generation, storage, export
- Layer 5: Perception Compiler — AI semantic extraction
- Layer 4: Template + Filter Engine — what and when
- Layer 3: Constitutional Layer — ethics, privacy, PII
- Layer 2: Skill Runtime — extensions
- Layer 1: HAL — hardware drivers (ESP32, RPi, cameras, sensors)

## Hardware Benchmarks

- GAP9 RISC-V: 18 fps, 62.9 mW, 9.3h battery (smart glasses)
- Ambarella CV72S: 4×5MP + AI, <3W (dashcams)
- Hailo-8L + RPi5: ~120 fps (batch=8), 4-5W, $70 AI Kit
- Snapdragon 8 Elite: ~45 TOPS, 56 models <5ms
- Full pipeline: capture(5ms)→detect(8ms)→classify(3ms)→filter(1ms)→spark(2ms) = 19ms → 52 fps

## Security (6 threat types)

- T1: Template Injection — fixed ontology, max 8 fields, no free text
- T2: Filter Surveillance — max 3 attributes, person-specific banned
- T3: Physical Prompt Injection — text quarantine, dual PII detection
- T4: Adversarial PII Bypass — cascade detection
- T5: LoRA/Model Poisoning — signed adapters only
- T6: Side-channel — Secure Boot, Flash Encryption, JTAG disable
- Resolution cap: 160×120 for always-on (facial recognition impossible)

## Applications

- Companion AI robots with emotional intelligence
- Autonomous driving training via consumer dashcams
- Business behavior analytics (retail, restaurant, clinic, construction)
- Smart glasses (solving the Google Glass privacy problem)
- Medical privacy (telemedicine without intimate photos)
- Automated scientific discovery (multimodal observation → physics laws)

## Self-Learning ("Grows With You")

AisthOS devices learn and evolve through three parallel tracks:

- Track A (FAST, real-time): Contextual bandits learn user preferences from reactions. Reflexion-style verbal self-analysis. Adapts in minutes.
- Track B (MEDIUM, nightly): Pattern mining across daily Sparks → LLM auto-generates new SKILL.md files. Device wakes up with new abilities.
- Track C (SLOW, weekly): MLX LoRA fine-tuning on accumulated Sparks. Adapts communication style and personality. 250-475 tok/s on Mac Mini M2.

Growth stages: Infant (days 0-3, basic reactions) → Child (weeks 1-2, pattern discovery) → Teen (months 1-2, self-created skills) → Adult (months 3+, unique personality).

User Wisdom is portable: learned preferences, skills, and personality can be exported and transferred to any AisthOS device. Create your companion once, use everywhere.

## Links

- Website: https://aisthos.dev
- GitHub: https://github.com/aisthos/aisthos
- License: MIT
- Author: Vladimir Desyatov
- Contact: GitHub Issues