Robotics and computer vision / 2024
Computer vision portfolios usually live in notebooks. Hard for hiring teams to actually run, and disconnected from real product surfaces. Streaming inference under backpressure is rarely shown end-to-end.
React/Vite client captures webcam or video frames, sends them through a WebSocket to a FastAPI frame processor, and renders bounding-box overlays plus throughput, latency, and confidence telemetry. A detector abstraction lets the system run with a deterministic mock detector by default and an optional YOLOv8 adapter when weights are supplied.
My part
Decisions
In real-time vision systems, processing a stale frame is often worse than skipping it. The world has already moved on. The pipeline acknowledges one frame at a time and drops anything that arrives during processing.
Reviewers can verify the streaming architecture, backpressure behaviour, and dashboard wiring without GPU drivers or model weights. The mock is honestly labelled and emits deterministic detections.
Maximum connection counts, target FPS limits, and maximum frame-pixel limits sit at the WebSocket boundary so a misbehaving client cannot saturate the worker pool.
SYSTEM ARCHITECTURE
Frame capture and pixel limits
One frame waits for one acknowledgement
Connections, FPS and payload bounds
Deterministic mock or optional YOLOv8
Boxes, confidence and live metrics
Fresh visual feedback
Built with
Links
More work
ARIRA ROS-based autonomous retail robot combining real-time shelf perception, multi-object tracking, autonomous navigation, customer-request handling, and ERP synchronisation.Validated in simulationRobot Guide DogReal-time ROS perception and audio-guidance system developed for a Unitree Go1 guide-dog concept, combining object detection, scene context, and synchronised spoken cues.Research prototypePath Planning SimulatorInteractive browser-based simulator comparing A*, Dijkstra, and an RRT-style planner on a configurable grid with obstacle painting, start/goal placement, and live stats.Maintained