Occupancy intelligence with ML-powered detection

Context

Knowing what capacity is actually in use, right now, across multiple sites was a manual exercise: staff walked the site, counted, and recorded the result. The number was already stale by the time it reached anyone who could act on it, and it disagreed with the reservation system often enough that nobody fully trusted either.

Capacity decisions and short-stay pricing were therefore made on judgement rather than data.

Challenge

Occupancy is deceptively hard. Physical presence and booked presence are different facts, and reconciling them requires both a reliable sensing method and a clear operational definition of what counts. The product problem was as much about definitions and exception handling as about detection accuracy.

The system also had to work across sites with different layouts, lighting and conditions — without a bespoke configuration effort per location.

Approach

Led development of an ML/AI-powered detection system to establish real-time physical occupancy, defining the accuracy thresholds, failure behaviour and human-override paths the operation could actually work with.

Built the occupancy-management feature around exceptions. Rather than presenting a raw feed, the product surfaces the discrepancies between expected and detected state — the small set of cases where someone needs to act — instead of asking staff to review everything.

Fed occupancy into resource allocation. Accurate live capacity meant short-stay availability could be released with confidence rather than held back as a safety margin.

Kept humans authoritative. Detection informs; operators confirm on exceptions. This was essential to trust, and it also generated the correction signal that improved the model over time.

Outcome

• 30% of daily manual work automated, removing the recurring physical count.
• Occupancy accuracy improved and reconciled against the reservation system.
• 10% uplift in transient revenue, by releasing capacity that had previously been withheld as buffer.
• Occupancy became a shared operational input rather than a contested figure.

What I would carry into another engagement

The value was not in the model's accuracy in isolation — it was in defining what the organisation would do differently at each level of confidence. An ML feature without an agreed decision it feeds is a demo; the same model wired to a specific operational choice is a revenue line.