Mitterrand
Ekole

End-to-end visual-inertial odometry system estimating 6-DoF pose in real-time by fusing stereo cameras with an IMU. Features both CPU baseline and GPU-accelerated pipeline using custom Metal compute kernels.

• IMU preintegration with bias Jacobians
• 6 custom Metal compute shaders
• GPU reduces worst-case latency by 1.8x

Safe Rust library wrapping Metal compute shaders for classical computer vision algorithms. Designed for onboard robotics pipelines with zero-copy and no heap allocations in hot paths.

• Zero unsafe code in user-facing API
• Compile-time GPU/CPU access safety via ownership
• 8 Metal compute shaders for CV primitives

Safe, ergonomic Rust bindings for Apple's MLX array framework, enabling ML workloads on Apple Silicon without Python. Includes optional Apple Neural Engine offload.

• MLP training 2.1x faster than Python MLX
• ANE inference 1.51x faster than GPU
• ~81,000 LOC across 87 Rust files

AI agent exploring proof spaces for open mathematical problems using Monte Carlo Tree Search with LLM-based proof generation and dual formal/informal verification.

• UCB1 node selection with geometric backpropagation
• Dual verification: Lean 4 formal + LLM informal
• Applied to Navier-Stokes problem exploration

3D scene reconstruction from unordered images with Transformer-based matching for robust multi-view geometry.

• Transformer-based feature matching
• Unordered image set reconstruction
• Multi-view geometry pipeline