Face recognition attendance for 2,000+ people, running entirely offline

The problem

More than 2,000 people moved across several institutes, and keeping track of them was manual work. Attendance had to be recorded and then reconciled, and any question after the fact — who was on site on a given day, was a particular area staffed — meant going back through records by hand.

At that scale, manual attendance also has a credibility problem. It depends on people recording things accurately and honestly, and it is difficult to stand behind when compliance rests on it.

What we built

A computer vision system that works off the camera infrastructure already in place, so there was no new hardware to roll out across multiple sites.

• Face recognition on live video streams — identifying people from existing feeds rather than at a dedicated checkpoint or turnstile.
• Automatic attendance marking — presence recorded as people move through the campus, with no queue and nothing to sign.
• Presence tracking across institutes — a current view of who is on site, and a record that can be checked later.
• Compliance reporting — an auditable record that does not depend on anyone filling in a register correctly.

Built in Python with OpenCV handling the video pipeline and MariaDB behind it, delivered over three to four months.

Privacy: it never leaves the building

The whole system runs on a server installed on site, with no internet connection. Face data is captured, processed and stored locally and never leaves the premises. There is no cloud service in the loop and no third-party processor holding anyone's biometrics.

For face recognition this is the question that decides the project. The usual objection is "where does our people's face data actually end up?" — and here the answer is simply: on your own server, in your own building, on a machine that isn't connected to the internet. It removes an entire category of risk rather than managing it.

Why the offline design was the right call

• No remote exposure. A system with no internet connection cannot be reached from outside it.
• It works when the connection doesn't. Attendance keeps running through outages and patchy campus connectivity, because nothing depends on an external service being up.
• No per-user cloud costs. Running on their own hardware means the bill doesn't grow every time headcount does.
• No hardware rollout. Reusing the cameras already installed kept the cost and disruption of covering multiple institutes down.

The results

Attendance for over 2,000 people is now marked automatically instead of being collected and reconciled by hand, which removed a recurring daily administrative task across every institute. Tracking who is on a campus stopped being a matter of asking around, and compliance rests on a system record rather than a paper trail that has to be trusted.

Built with