ElMonstroDelBrest/ChaosAI
Four-stage deep learning pipeline: Spherical VQ-VAE tokenizer → Mamba-2 JEPA → Stochastic multiverse predictor → Latent regime RL agent. Trained on H100.
What's novel
Four-stage deep learning pipeline: Spherical VQ-VAE tokenizer → Mamba-2 JEPA → Stochastic multiverse predictor → Latent regime RL agent. Trained on H100.
Code Analysis
12 files read · 4 roundsA self-supervised Mamba-3 JEPA foundation model for financial time series that uses CFM regularization to eliminate evaluation leakage, with a 4-stratum pipeline (FSQ tokenizer → JEPA encoder → QR-DQN → Multiverse Crossing) for crypto portfolio allocation.
Strengths
Genuinely novel architecture (Mamba-3 with complex RoPE, trapezoidal discretization, CFM-as-regularization, geodesic Multiverse Crossing) with mathematically rigorous implementation and clean JAX/Flax functional style. The leakage measurement finding and CFM elimination mechanism are substantive research contributions, not just engineering.
Weaknesses
Test coverage is limited (10 test files for 166 source files) and I couldn't verify their depth. The project mixes PyTorch and JAX ecosystems, creating dependency complexity, and some config defaults (d_model=1024, n_layers=24) don't match the 33M parameter model in the README, suggesting configuration fragmentation.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
69
Contributors
1
Files
65
Active weeks
3
Repository
Language
Python
Stars
2
Forks
0
License
AGPL-3.0