What wasn't working.
A large industrial site relied on thumbprint biometric attendance for nearly 30,000 workers. During every shift change, eight parallel biometric lanes operated for almost two hours, creating long queues, lost productivity, and significant hardware maintenance overhead.
The existing approach was difficult to scale. Adding more fingerprint devices would only increase infrastructure complexity without solving the underlying bottleneck.
During COVID-19, the same challenge appeared in healthcare environments. Hospitals needed to replace shared fingerprint devices because they had become a hygiene risk, and doctors and nurses lost time walking to dedicated entry points instead of checking in from their own departments.
The challenge was clear: build a secure, distributed, and touchless attendance system that could work across different environments.
What we actually did.
As Technical Project Manager and ML Engineer, I led a 12-member team across computer vision, machine learning, web development, QA, DevOps, and support to build an AI-powered facial recognition attendance platform.
Instead of scaling fingerprint infrastructure, we replaced it with a camera-based recognition system. Live video streams from IP cameras were processed in real time, detecting multiple faces in a single frame and matching employees against the workforce roster in under a second.
For industrial environments, the system reduced attendance bottlenecks by moving from eight biometric lanes to two facial recognition lanes. Attendance data from multiple sites was synchronized with a central platform and integrated with ERP systems for automated payroll and reporting.
The same AI engine was later adapted for healthcare and corporate environments during COVID-19. We built touchless check-in solutions using Raspberry Pi and Android tablets with cameras, so employees could authenticate remotely without touching shared devices. The platform integrated with ERP systems and turnstiles to support distributed access control and attendance management.
Where the interesting calls were.
The biggest decision was moving away from traditional biometric hardware instead of improving the existing setup. Adding more fingerprint readers would have delivered a short-term improvement but increased maintenance costs and preserved the same single-point bottleneck.
We chose an AI-powered, distributed architecture that required more upfront investment in computer vision and infrastructure design. The tradeoff was a more scalable system that reduced physical dependencies and could adapt across industrial, healthcare, and corporate environments.
Another challenge was handling recognition failures. Even at 98% accuracy, edge cases needed a separate flow. We designed exception handling so unsuccessful matches did not slow down the main process or affect everyone else.
The numbers after we shipped.
The system reduced industrial shift changeover time from 2 hours to 30 minutes and cut attendance queues from eight lanes to two. It reached 98% recognition accuracy across a workforce of more than 30,000 employees.
During COVID-19, the same platform enabled touchless attendance and access control in healthcare environments, removing shared fingerprint touchpoints and letting staff check in closer to their actual workplace.
The solution evolved from an attendance system into a flexible workforce management platform integrated with ERP systems, payroll, and access control infrastructure.
Things I'd carry into the next case.
At enterprise scale, the biggest bottlenecks are often created by physical processes rather than software limitations. Replacing a slow process with AI is not only about improving accuracy, it is about redesigning the workflow around how people actually work.
The last 2% of AI accuracy matters as much as the first 98%. A production AI system succeeds only when it has a clear strategy for handling exceptions without disrupting the experience for everyone else.
Designing for real-world constraints, hardware limitations, connectivity, and data privacy from the beginning creates solutions that are more practical and resilient than technology-first approaches.