Catching what CCTV couldn’t

The challenge

Our client had reached near-profitability, but persistent inventory losses across hundreds of warehouses were quietly eroding their margins. Employee theft, mostly low-value items, was happening consistently across the network: small per-incident, significant at scale. Existing approaches, including strict operational processes, frequent stock counts, and manual CCTV review by a small fraud team, couldn’t cover that many locations. The opportunity was to build something that could process footage automatically and surface only the cases worth acting on. That started with reframing the problem from “detect theft” to “identify specific, machine-detectable behaviors that indicate suspicious activity.”

Our approach

Before writing any code, we evaluated existing market solutions. None could meet the required level of customization and integration, so we built our own. The work ran in one-week cycles with frequent demos and proactive updates to stakeholders, pushing information out continuously rather than waiting for approvals. The client’s fraud team helped calibrate what actually looked suspicious in a warehouse context, and that domain knowledge proved as important as any technical decision. We pushed back on internal processes slowing progress, on infrastructure assumptions that didn’t hold up early, and on the instinct to wait for a perfect solution before shipping anything.

The result

We delivered a vision AI anti-theft system that continuously monitors warehouse activity through existing CCTV infrastructure, detects suspicious behavior using computer vision, correlates it with inventory data, and surfaces flagged cases to the fraud team via Slack and a dedicated web interface. Privacy-preserving masked video ensures full legal compliance. The fraud team now reviews only high-signal, pre-qualified cases instead of hours of raw footage. Lumen is live in a pilot set of warehouses where it generates real detection data, with financial impact metrics still coming in from the rollout.

Why this matters to us

This project is a good example of what happens when you treat a complex AI problem as a product problem first. The technology only works if you understand the domain, the users, and the real-world constraints of the environment it has to operate in. Getting that right is the work we love, and where we think we do our best.