Service · Computer vision & edge AI

Computer vision that works live, at the edge.

Q: Should vision run on the edge or in the cloud?

If you need low latency, can't upload high-resolution video, or must keep working when the connection drops, run it on the edge. Cloud processing suits batch analysis where a delay is fine. Many systems combine both.

Q: Which hardware do you work with?

Raspberry Pi 5 with camera modules such as the Arducam IMX708 for prototypes, and NVIDIA Jetson for heavier, multi-camera workloads.

Q: Can you work with our existing cameras or footage?

Usually, yes. We start by testing on your real footage to see what accuracy is achievable before committing to hardware changes.

Q: Do you also handle the hardware and installation side?

Yes — camera selection, compute, wiring, enclosures and mounting are part of how I deliver vision systems.

I build real-time vision systems that run next to the camera — detection, tracking and the decisions on top of them. For Urban Soccer Park I built an AI camera that tracks every player and the ball and directs its own virtual camera, with all the vision running on edge hardware at the pitch.

Book a 20-min call See the case study

6+ yrs: in AI & computer vision
Edge: inference — no cloud round-trip in the loop
~150°: wide feed reframed into a broadcast view

01 — Is this you?

Problems I'm usually brought in to solve.

Decisions must be real time

Live video needs answers in milliseconds, not after an upload.

Cloud isn't an option

Bandwidth, privacy or unreliable links mean processing has to happen on the device.

Hard visual conditions

Small, fast or often-occluded objects that off-the-shelf models miss.

Hardware is part of the problem

You need help choosing cameras and compute, not just training a model.

02 — What I build

What you get.

Object detection & tracking

Detection and multi-object tracking that keeps identities stable through occlusion.

Edge deployment

Models optimised to run on Raspberry Pi and NVIDIA Jetson class hardware.

Camera & hardware design

Camera modules, MIPI CSI, compute, enclosures and mounting — tested on site.

Decision & control logic

Turning detections into actions — e.g. Kalman filtering and PID control for smooth camera motion.

Video pipelines

Capture, inference and output as RTSP streams and recordings.

Performance tuning

Latency budgets and graceful failure, so the system keeps running in the real world.

03 — Proof

CASE STUDY

REAL-TIME · ON-DEVICE

AI-Powered Soccer Camera System

A physical AI system that perceives live soccer, tracks players and the ball, and produces a broadcast-style view with no camera operator — running on edge hardware.

Read the case study →

04 — How we'll work

01
Discovery call
We talk through the problem, your data and constraints, and what success looks like. You get an honest view of what AI can and can't do here.
02
Architecture & plan
A clear design — components, models, data flow and how quality will be measured — with a scoped first version.
03
Build & evaluate
Working software early, tested against real examples from your domain rather than demos.
04
Ship & improve
Deployed to production, monitored, and tuned for accuracy, latency and cost as real usage comes in.

Technology

Python

PyTorch

YOLO

OpenCV

GStreamer

CUDA

RTSP

Kalman / PID

Raspberry Pi 5

NVIDIA Jetson

MIPI CSI

05 — FAQ