Clients wanted to know whether cleaning was on track without waiting for end-of-day updates, manual reports, or calls to the operations team.
Admins had to decide manpower allocation across sites with incomplete context. A site might need more cleaners because of higher workload, repeated missed tasks, event-driven demand, or quality issues, but those signals were spread across schedules, checklists, attendance, inventory, and site reports.
Cleaners were already uploading photos as proof of work, but reviewing those photos manually did not scale. Supervisors needed a way to spot-check whether work looked up to standard without inspecting every image themselves.
We began by connecting a practical slice of live operations data: scheduled tasks, submitted checklists, cleaner attendance, uploaded photos, and site status.
On top of that data, we prototyped a client-facing AI agent that could answer operational questions such as what had been completed, what was still pending, which areas had photo evidence, and whether any site needed attention.
We paired that with a multimodal photo triage agent. Uploaded photos were reviewed against expected cleaning standards, with low-confidence or likely-failed checks escalated to admins instead of buried inside a gallery.
The production system connected iOS, Android, web, and Telegram to the same backend. Cleaners could check in, submit checklists, upload proof-of-work photos, update task progress, and raise operational issues from the surface they already used.
A real-time client agent sits on top of the cleaning management data. Instead of waiting for a report, clients can ask what happened at their site today, which areas are complete, which tasks are delayed, and where photo evidence exists.
An admin planning agent reviews workload, attendance, open issues, site history, cleaning frequency, and quality signals to advise whether a site needs more manpower, a schedule change, or supervisor attention.
A multimodal QA agent triages cleaner-uploaded photos. It checks for visible cleanliness signals, missing proof, poor photo quality, and likely non-compliance, then sends uncertain or failed cases into a supervisor review queue.
Underneath the agents is the full cleaning management system: site setup, recurring schedules, manpower assignment, inventory, ticketing, checklists, attendance, evidence storage, dashboards, and exportable client reports.
- Clients gained real-time answers about cleaning progress instead of waiting for manual updates.
- Admins gained AI-assisted manpower recommendations grounded in live operational data.
- Photo evidence became an active quality-control signal, not just an archive for later reporting.