What makes a video analytics project succeed?
Designing what happens after each alert: who receives it, what they are expected to do and by when, who hears if it isn't done, and how repeat problems get fixed at the root. Accuracy, speed and uptime are necessary, but an accurate alert that reaches the wrong person, or no one, changes nothing.
A perfectly accurate alert that lands in a group chat of forty people, or on a dashboard nobody opens, has the same effect as no alert at all. Projects with excellent models fail this way, and projects with ordinary ones succeed, because in the second kind every alert has an owner and a next step.
1. Send it to one person who can act
Every alert type needs an owner: the person nearest the problem, with the authority to fix it. A table left uncleaned belongs to the floor staff on shift, not the regional manager. A forklift in a walkway belongs to the shift supervisor. Send everything to everyone and everyone assumes someone else has it.
- Route by alert type and by site, not to one big group.
- Send it where people already are. For most store and floor teams, that is WhatsApp.
- Put the photo, the place and the time in the message, so nobody has to log in to understand it.
2. Say what is expected, and by when
An alert should carry an obvious next step. “Table 6 not cleaned” means clean table 6. “Back door open after 10 pm” means close it and find out why. When the expected action and the time allowed are agreed in advance, the alert becomes an instruction rather than a notification.
3. Decide who hears if nothing happens
Some alerts will be missed. The escalation path decides what happens next. For example: the floor staff gets the alert; if the problem is still there after 15 minutes, the shift manager hears; if it happens every lunch hour for a week, it shows up in the area manager’s weekly review. Each step up should be rarer and more serious, so senior people see patterns, not noise.
4. Treat repeat problems as a process problem
An alert that fires at the same place every day is no longer about the staff. It is telling you something about the process: too few people at lunch, a bin that is too small, a door that doesn’t close by itself. Chasing every occurrence wastes everyone’s time. Look at the pattern once, fix the cause, and the alerts stop.
5. Use 80-20 to decide what matters
In most businesses, a handful of checks produce most of the issues, and a handful of sites produce most of those. Start there. A short list of three things to fix today, for each site, gets acted on. A dashboard of forty metrics gets admired. Add checks only once the first ones are under control.
Accuracy still matters, for a different reason
None of this makes accuracy unimportant. False alerts are what kill a workflow: after a few wrong ones, people stop reading. That is why every Tarsyer alert is checked by a second, larger model before it is sent, and anyone can mark a wrong one with one tap. Accuracy earns attention. The workflow turns attention into results.
Where Tarsyer fits
When we set up a site, we agree the workflow before the checks: who gets each alert, what they should do, how it escalates and what goes into the weekly review. Store managers can start the day with a short list of things to fix, and head office sees which problems keep coming back.