AI for Advertising & AdTech
Digital out-of-home has spent a decade selling impressions it could not measure and scheduling campaigns by clock rather than audience. Computer vision changes the input; the hard part is doing it without touching anyone's identity.
Static playlists on expensive inventory
Conventional signage runs predefined playlists on time-based schedules, so a campaign plays to whoever happens to be standing there. Targeting is a manual scheduling exercise, engagement drops through every mismatched audience period, and advertisers get almost no real-time insight into who actually saw a spot. The instinct is to fix it with facial recognition, which is the wrong answer commercially as well as legally: it converts an audience-measurement problem into a biometric-identification problem, with everything that follows. The useful signal is contextual rather than personal, and getting that distinction right in the architecture is what makes the rest of the system shippable.
What AI actually does in advertising & adtech.
The four workloads that carry the value in this industry, in the order they usually pay off.
Anonymous audience analysis
Cameras capture the scene and vision models extract contextual signals without identifying anyone: estimated audience size, general age bands, presence of adults or children, crowd density and time of day. No facial identities and no personally identifiable information are stored. This is the platform we built for smart digital billboards.
Real-time ad selection
A decision engine evaluates live context against active campaign rules and selects the most relevant creative instantly, rather than waiting for the next scheduled slot. Interactions and campaign performance are logged for continuous optimisation, which is also the reporting layer advertisers have been missing.
Knowledge-driven campaign rules
A retrieval layer lets the platform pull campaign rules, advertiser objectives, product information, seasonal promotions and brand guidelines before ranking eligible ads. Marketing teams update campaign knowledge as content rather than as code, so the logic governing which ad runs when changes without a redeploy.
Fleet operations across sites
One screen is a demo, a network is an operations problem. Remote model updates, per-device health monitoring, detection of a camera knocked out of alignment or obscured, and rollback without dispatching a van. The edge fleet has to be treated as infrastructure or the rollout becomes a permanent firefight.
The privacy boundary is an architecture decision, not a policy page
Under the GDPR, biometric data processed for the purpose of uniquely identifying a person is special category data and carries a materially heavier compliance burden. Anonymous audience measurement that estimates aggregate characteristics without identifying or re-identifying individuals sits on the other side of that line, and staying there is a design constraint rather than a promise: no identity template is generated, no face image is retained, and the signals leaving the device are aggregate. The EU AI Act tightens the surrounding area further, restricting real-time remote biometric identification in publicly accessible spaces and prohibiting emotion inference in workplace and education contexts. We scope where a system sits relative to those lines before any camera is specified, because the answer determines the hardware, the retention policy and whether the deployment is viable at all.
The practice areas this work draws on.
Advertising & AdTech questions, answered.
No, and that is deliberate. The system performs anonymous audience analysis: aggregate signals like estimated audience size, general age bands and crowd density. No facial identities and no personally identifiable information are stored, which keeps the deployment out of the special-category biometric regime rather than inside it with mitigations bolted on.
Yes. Campaign rules, advertiser objectives, product information and brand guidelines are retrieved as content rather than compiled into application logic, so updating them is an editorial action. That is the main reason this architecture survives contact with a real campaign calendar.
A camera and enough on-device compute to run inference locally, which is also what keeps raw imagery from leaving the site. The specification depends on the signals you need and the environment the screen sits in; we size it during scoping rather than quoting a single reference build.
Interactions and campaign performance are logged per play, so reporting is built from aggregate exposure and context at the moment of display rather than modelled from footfall estimates. It is measurement rather than identification, and that distinction is what makes it reportable.