Public housing estate construction in Singapore is one of the most safety-regulated construction environments in the world. This case study looks at what happened when AI video analytics began watching an HDB estate build through the CCTV already on site, around the clock, and feeding what it saw back to the people who could act.
CLIENT: HDB SINGAPOREAn estate build packs an unusual amount of simultaneous risk into a tight footprint. Structural work, M&E installation, wet trades and finishing crews operate above and beside one another, and the picture changes week by week as blocks rise and handover dates approach. Every trade brings its own supervisors, its own method statements and its own habits.
Safety teams on projects like this are rarely short of rules. They are short of eyes. A supervisor doing rounds sees each work front for minutes at a time, then moves on. Manual observation programmes produce a steady trickle of cards, but they record what people happen to notice, and people tend to notice the same things in the same places. Between inspections, compliance drifts. Not because anyone intends it to, but because production pressure is constant and human attention is not.
None of this is specific to one project or one contractor. It is the operating reality estate construction teams manage everywhere, and it does not respond to more paperwork. It responds to being seen.
SafeKey, the AI detection layer of the Invigilo platform, connected to the CCTV already installed on site. Any camera with an RTSP stream can feed the system, so a standard deployment needs no new cameras and no new cabling, and a site is live within 24 hours of camera access.
Zone-based rules pointed the AI at the behaviours that hurt people on estate builds: work at height, PPE compliance and entry into restricted zones. The platform draws on 60+ computer vision models covering 40+ risk types, with 85%+ detection accuracy verified on live site footage rather than benchmark datasets. When a rule is breached, the alert reaches the site team in under a second on WhatsApp, Microsoft Teams, Telegram or email, following an escalation chain the team defines.
Every detection is saved as a timestamped clip. That single design decision matters more than it sounds, because it turns each alert into evidence a supervisor can act on and a toolbox talk can be built around. The system flags conditions, not identities: there is no facial recognition, deployments run edge, on-premise or hybrid, and the platform carries long-standing ISO 27001 certification and is built for GDPR and PDPA compliance.
Connects to the CCTV a site already owns. Any RTSP stream works, so there is no new hardware for a standard deployment and no waiting on procurement.
Each camera view is divided into zones with their own rules, so an edge zone can demand harness discipline while a walkway watches for restricted entry.
WhatsApp, Microsoft Teams, Telegram and email, in under a second, routed through escalation chains so the right person hears first.
Every detection becomes a timestamped clip. Nothing depends on someone remembering to write down what they saw.
The results are connected. The observation volume came first, the response speed followed from it, and the incident reduction is what the two of them produced together.
The first change was volume. A patrol sees a work front for minutes; the cameras watch it for every hour work happens, including the hours no round was ever going to cover. Safety observation stopped being a sampling exercise and became a continuous record. The team went from reconstructing what probably happened to seeing what actually happens, in the corners of the site that footpaths and schedules rarely reach.
Speed followed from routing. When a detection fires, the alert lands on the phone of the person who can act, not in a report that surfaces at the next meeting. The distance between something going wrong and someone knowing about it shrank from the length of a reporting cycle to seconds, which means a supervisor can intervene while the condition still exists rather than investigate it after the fact.
Fewer incidents is the outcome, not the mechanism. When unsafe conditions are visible every time they occur, and the response arrives while the work is still in front of the worker, behaviour shifts. Risky shortcuts stop being invisible, corrections happen in minutes instead of days, and the incident count falls as a consequence. That is where the 60% reduction came from.
A dashboard full of alerts that nobody acts on is just a record of a site getting hurt. What made this deployment work is that the observations flowed into routines the site already ran. Timestamped clips gave toolbox talks something concrete to show instead of a generic reminder, and corrective actions could be assigned against evidence rather than recollection. Each detection fed the next conversation, and each conversation made the next detection rarer.
Workforce trust carried the rest. Because the system flags conditions rather than identities, with no facial recognition anywhere in the stack, the conversation on the ground stayed about the work and not about surveillance. That is what let the loop keep closing, week after week, without the programme turning into a policing exercise.
A proof of concept on your own CCTV, live within 24 hours of camera access.