| Repo |
Focus |
What you can check |
| Boundaryd |
Rust policy gateway |
Offline-capable decide API with PostGIS geofences, hash-chained audit, transactional outbox; 2.53M evals/s in-process; cached HTTP 15.3k rps / p50 0.99 ms on a Ryzen 5 3600 |
| Staylock |
Java 21 inventory |
Hotel reservation ledger with a last-room oracle: 200 concurrent guests → 1 hold / 199 sold out / 0 remaining, asserted in CI; explainable ranking; hold-then-confirm API and public demo UI |

| Repo |
Focus |
What you can check |
| FieldCast |
Rust + GStreamer |
Real-time two-camera composition and H.264/MP4 output; custom zero-copy continuity element, structured timestamp diagnostics, clean-EOS drain, real media smoke test in CI |
| FrameForge VK |
C++20 + Vulkan compute |
Chained grayscale/Sobel shaders with explicit barriers and timestamp queries; 1.67 ms at 1080p on RX 5600 XT vs 3.76 ms CPU (2.25x), max error below 3e-7 |
| TrailVision |
Multimodal sports video |
Auditable voice/IMU/vision event fusion, SQLite replay, FastAPI, and responsive review UI; deterministic demo fuses 17 candidates → 6 events → 3 runs |
Accepted upstream:
| Project |
Change |
Verification |
| ROS Controls #2567 |
Adds closed-form bicycle, tricycle, and Ackermann integration tests |
Maintainer-approved and merged; exact Rolling package passed 32/32 tests |
Selected patches currently under maintainer review:
| Project |
Change |
Verification |
| Gazebo Sensors #640 |
Bounds reused thermal and segmentation message headers instead of appending one entry per frame |
Full local suite; Homebrew ARM64, Ubuntu, and Windows hosted builds green |
| ros2-rust #669 |
Coalesces duplicate executor task wakes while preserving wake-during-poll |
Nine hosted builds green across Humble, Jazzy, Kilted, and Rolling |
| RViz #1837 |
Surfaces incompatible QoS as a visible error and recovers after compatible traffic arrives |
Fast DDS reproduction; exact core and downstream display checks pass locally |
IROS 2026 Indoor Exploration Competition — CPU-only policy. On a released single-robot map: 65.17% base-station coverage vs 35.02% nearest-frontier baseline in matched local runs. Cross-map and hidden-start robustness testing is in progress.
CS, Trinity College Dublin.