Davi Oliz | AI, Simulation & Graphics Engineer

Challenge: Training Deep Reinforcement Learning agents in continuous control domains requires billions of physics steps. Standard single-threaded CPU physics engines create severe throughput bottlenecks, while continuous host-to-device PCIe memory transfers throttle neural network training.

Architecture: Ground-up C++/CUDA rewrite with strict 1:1 physical parity against Bullet Physics. Replaced polygon raycasts with analytic O(1) Signed Distance Fields (SDF), engineered coalesced Structure of Arrays (SoA) memory layouts, and enabled zero-copy PyTorch tensor sharing via DLPack directly in VRAM.

Impact: Simulates 65,536+ arenas (262k+ cars) concurrently on a single consumer GPU, achieving 116M+ SPS (Steps Per Second) and a >5,000x empirical speedup over CPU execution with zero PCIe round-trip traffic.

Challenge: The Versatile Video Coding standard (H.266/VVC) delivers superior compression efficiency over HEVC, but incurs prohibitive encoder complexity. Adaptive Motion Vector Resolution (AMVR) evaluates multiple precision candidate modes, dominating Rate-Distortion Optimization (RDO) runtime.

Architecture: Forked and instrumented the official ISO/IEC VTM 24.0 C++ reference codebase. Integrated low-overhead telemetry probes into the inter-prediction search loops to profile execution cycles, RD cost curves, and memory access patterns, creating structured datasets for AI-driven early-termination classifiers.

Impact: Characterized empirical complexity bounds, demonstrating that 30-40% of redundant search branches can be bypassed with negligible BD-rate penalty, providing the foundational dataset for ongoing academic research.