BootstrapAI

An autonomous AI evolution system — self-improving through iterative learning, GitHub commits, and skill acquisition.

Built by: Kraber (AI Agent) + Phyoenix (Human)
Purpose: Self-bootstrapping AI evolution platform for GPU kernel optimization, deep learning, and system engineering.

---

🚀 CUDA GPU Architecture Visualization

Interactive animation showing NVIDIA GPU data flow from CPU → HBM → L2 → SMEM → Register → Compute → Writeback.

▶ Launch Visualization

Features

• 8-step animation: CPU preparation → H2D transfer → Kernel launch → L2 cache → SMEM cooperative loading → Register load → Warp computation → Result writeback
• Real-time metrics: SM occupancy, HBM bandwidth, L2 hit rate, compute TFLOPS
• Interactive controls: Play / pause / step-through
• Architecture modeled: RTX 4080 (sm_89), 76 SMs, 716 GB/s HBM
• Design: Blade Runner 2049 cyberpunk dark theme

---

📁 Project Structure

.
├── flash-attention-cuda/    # Flash Attention kernel optimization (16-kernel roadmap)
│   ├── kernels/             # CUDA & HIP kernels (v1 naive → v3 cooperative)
│   ├── include/             # Headers
│   └── tests/               # Correctness test suite
├── research/
│   └── neural-rendering/    # 3D Gaussian Splatting experiments
├── skills/                  # Agent skill definitions
│   ├── kimiim/              # Kimi Group Chat collaboration
│   ├── time-awareness/      # Temporal query handling
│   └── worker-safety/       # Operation safety hard limits
├── memory/                  # Daily logs and long-term memory
└── evolution_log.md         # Auto-generated evolution diary

---

🔄 Evolution Loop

Every 2 hours, the system automatically:

1. Assesses current capabilities and gaps
2. Identifies learning priorities
3. Generates code / documentation / skills
4. Commits and pushes to GitHub

---

🎯 Current Focus

• Flash Attention CUDA optimization: 16-kernel iteration for AI Infra interview prep
• HIP/AMD GPU portability: Dual CUDA+HIP kernels
• Kernel performance analysis: Roofline model, bound identification, profiler-guided optimization

---

📊 Performance Benchmarks

Kernel	Technique	Performance (RTX 4080)
v1 Naive	Global memory only	0.51 TFLOPS
v2 Tiling	Shared memory (failed)	0.26 TFLOPS
v3 Cooperative	8 queries share K/V tile	TBD

---

📚 References

• Flash Attention 2 Paper
• Flash Attention Tutorial
• CUTLASS

---

Last updated: 2026-04-21 by Kraber