Senior Computer Vision Engineer

Lexerus; United Kingdom; Remote; GBP 70,000 - 110,000; Full time

Job summary

Lexerus is a UK-based retail‑tech company building AI-powered loss prevention for supermarkets and independent retailers. As a Computer Vision Engineer, you’ll own the detection stack—from person detection to behaviour classification—running on modest in-store hardware, with an emphasis on accuracy and cost.

You’ll fine‑tune models on real footage, drive down false positives, and optimize inference for edge deployment.

Qualifications

• 4+ years building computer vision systems for production.
• Strong Python skills with practical PyTorch experience.
• Hands-on in object detection and multi-object tracking (YOLO-family, ByteTrack).

Responsibilities

• Own the end-to-end detection pipeline: person detection, multi-object tracking, re-identification, behaviour classification.
• Fine-tune and evaluate models on real store footage and build evaluation sets.
• Reduce false positives to keep retailers engaged with the system.
• Optimize inference for edge hardware, balancing accuracy, frame rate and cost.
• Build active-learning loop turning user feedback into new training data.
• Collaborate with pilot stores to understand misses and false alarms.

Skills

4+ years production-ready CV

Strong Python

Object detection & multi-object-trt

Tools

PyTorch

OpenCV

YOLO-family

ByteTrack

Job description

Lexerus is an early‑stage retail‑tech company building AI‑powered loss prevention for UK supermarkets and independent retailers.

Our platform watches live camera feeds and flags shoplifting as it happens — not hours later in a review queue. Detection runs on‑device in the store, so footage never leaves the premises unless a retailer chooses to keep it as evidence.

We're deployed in live stores today, and we're scaling the product, the detection models, and the team.

The Impact You’ll Make

As our Computer Vision Engineer, you'll own the part of Lexerus that everything else depends on: whether we actually catch theft, and whether we cry wolf while doing it.

You'll work on the full detection stack — person detection and tracking, re‑identification across occlusions, behaviour classification, and the suspicion scoring that decides whether a shop manager's phone buzzes. It all runs locally on modest store hardware, so accuracy and inference cost are the same conversation.

This is a hands‑on role for someone who wants their models in production under real conditions — bad lighting, awkward camera angles, crowded aisles — rather than on a benchmark leaderboard.

What You’ll Do

• Own the detection pipeline end‑to‑end: person detection, multi‑object tracking, re‑identification, and behaviour classification.


• Fine‑tune and evaluate models against real store footage, and build the evaluation sets that tell us whether a change actually helped.


• Drive down false positives — the single metric that decides whether a retailer keeps the system switched on.


• Optimise inference to run reliably on in‑store hardware, balancing accuracy against frame rate and cost.


• Build the active‑learning loop that turns user feedback on each alert into the next round of training data.


• Work directly with pilot stores to understand why a miss or a false alarm happened, and close the gap.

Who You Are

• 4+ years building computer vision systems, with models you've taken to production and maintained there.


• Strong Python, and deep practical experience with PyTorch and OpenCV.


• Hands‑on with object detection and multi‑object tracking (YOLO‑family, ByteTrack or similar), including the unglamorous parts — data collection, labelling, and evaluation.


• Comfortable reasoning about inference performance: quantisation, batching, and what actually runs on constrained hardware.


• Rigorous about measurement. You don't ship a model change without evidence it beat the previous one.


• Pragmatic and product‑minded, with a bias for shipping something real over perfecting something theoretical.

Bonus Points For

• Experience with person re‑identification or metric learning (DINOv2, CLIP, or similar embedding models).


• Worked with vision‑language models for behaviour or scene understanding.


• Background in video analytics, CCTV, or RTSP stream handling at scale.


• Experience deploying and supporting on‑premises or edge systems.


• Prior startup or 0-1 experience.

Why Lexerus

Real‑world impact: UK retailers lose £4.2bn a year to theft. The people you're building for are shop staff who have largely stopped reporting it because nothing happens when they do.

Hard problems, real constraints: Real‑time computer vision on modest in‑store hardware, running all day, with false positives that cost a retailer their trust. There is nowhere to hide a sloppy system.

Ownership from day one: We're a small team. You'll own whole surfaces of the product, make the technical calls on them, and see your work running in a store the same week.

Short feedback loops: We talk to the shop managers using Lexerus directly. Feedback arrives in days, not quarters, and it shapes what we build next.

Early‑stage opportunity: Competitive salary, flexible remote working across the UK, and meaningful equity for early team members.