Projects | Owusu Kantankah

01open source · C++ · 3D transforms · animation systems

upstream merged Godot Engine

Godot Engine — Animation Regression Fix

Authored an upstream-merged fix for a Godot 4.7 animation regression by reproducing it across releases, using git bisect to isolate the offending change, and tracing it through the C++ animation pipeline.

root cause

A non-commutative matrix-composition error applied scale in the wrong coordinate space when position, rotation, and scale tracks were combined.

Identified Basis::scaled() as applying parent/global-space scaling where animated Node3D scale tracks required local-space semantics.
Corrected the composition with scaled_local(), preserving non-uniform scale with the expected R × S transform order during TRS animation.
Validated the regression across Godot releases and traced the behavior change back to the introducing engine optimization before submitting the upstream fix.

View Regression Issue Godot Repository

Godot animation regression and transform fix

C++ engine debugginglocal-space TRS scaling

02ML systems · research · graphics

Multi-Agent Reinforcement Learning System

An end-to-end PyTorch/CUDA training system for multi-agent policy experiments, paired with a browser visualization that executes exported policy weights in JavaScript.

2.3×training speedup

100agents in training system

40agents in current live demo

Batched per-group linear operations with torch.bmm, then used torch.compile and CUDA graph capture to reduce launch overhead.
Compared multiple policy architectures while keeping the environment and training setup controlled.
Reimplemented trained-policy inference for the browser and rendered autonomous agents in Three.js.

Launch Live Demo View GitHub Page

03computer graphics · performance

3D Drone Swarm

A real-time 1,000-agent swarm simulation implementing Reynolds-style flocking behavior in Three.js and GLSL.

133FPS after optimization

1InstancedMesh draw call

9 MBGPU memory

Collapsed 1,000 individual draw calls into one instanced dispatch.
Replaced O(n²) pairwise neighbor search with a 27-cell spatial-hash lookup.
Kept the physics loop allocation-free while preserving real-time flock behavior.

View GitHub Page

04C++ · rendering

CPU Raytracer

A C++ ray tracer built from first principles to understand the geometry and math underneath a rendering pipeline.

Implemented ray–sphere and ray–triangle intersections, surface normals, and camera projection.
Added diffuse, metal, and dielectric materials with reflections and refractions.
Supported antialiasing, depth-of-field, configurable scenes, and lighting.

View GitHub Page

05backend · deployment · teamwork

Centralized Morgridge Event Calendar

The backend and deployment layer for a centralized event calendar serving UW–Madison's CDIS community.

Built a RESTful Flask/SQLite API for event retrieval and channel management.
Deployed the data layer on AWS EC2 between automated scrapers and a React frontend.
Worked on a nine-person team and helped define the API schema and deployment pipeline.

View GitHub Page

06reinforcement learning

Flappy Bird Decision Boundary

A smaller RL experiment that made model behavior visible: a decision boundary from neural policies trained on Flappy Bird using REINFORCE and A2C.

I keep this one on the site because it captures something I still care about: models are more interesting when you can inspect what they learned.

Decision boundary plot from a Flappy Bird reinforcement learning experiment