Perfect-os

An AI-optimizable system playground. Not a rigid architecture — a playground of freedom.

The Problem

Modern computing wastes enormous resources:

• 75-95% of library code is never executed
• Syscalls are 10x slower than function calls
• Context switching burns up to 28% CPU in dense workloads
• io_uring proves 3-5x I/O throughput improvement is possible
• No OS applies profile-guided optimization system-wide

As AI models get smarter, the bottleneck isn't intelligence — it's system overhead. Disk access, bloated libraries, unnecessary services, unoptimized abstractions at every layer.

The Vision

Give an AI agent maximum freedom to observe, analyze, and optimize a Linux system at every level of abstraction. Bottom-up. From kernel parameters to userspace to applications.

┌─────────────────────────────────────────────┐
│                 AI Agent                     │
│          (the intelligence layer)            │
├─────────┬───────────┬───────────┬───────────┤
│ Observe │ Optimize  │   Bench   │  Safety   │
│  (eyes) │  (hands)  │(scorecard)│(safety net)│
├─────────┴───────────┴───────────┴───────────┤
│              Linux System                    │
│  kernel · services · packages · network · io │
└─────────────────────────────────────────────┘

Quickstart

# See what the system looks like
./perfect observe

# Run benchmarks
./perfect bench quick

# Save current state (safety first)
./perfect snapshot create "baseline"

# Find optimization opportunities
./perfect optimize scan

# Let the agent loop run (dry-run first)
./perfect agent --dry-run

# Let it optimize for real
./perfect agent --iterations 5

Architecture

observe/     → System profiling, waste detection, metrics collection
optimize/    → Kernel, services, packages, network, memory tuning
bench/       → CPU, I/O, memory, network benchmarks
safety/      → Snapshot and rollback system
agent/       → Observe → Optimize → Measure → Repeat loop
config/      → Baselines and optimization logs

Every module is independently useful. Zero external dependencies. JSON everywhere. Safe by default.

Philosophy

• Playground, not prison — the AI has freedom to explore and reshape
• Observe before you touch — always measure first
• Safe by default — auto-snapshot before every optimization, one-command rollback
• Zero deps — Python stdlib + system commands only
• Bottom-up — start from the kernel, work up through every layer