On-Device AI Safety System
A battery-powered device that warns children and cars of each other's presence. Built on edge AI under real constraints.
Situation
Incredilab set out to build a device that could warn children and cars of each other's presence. Real time, on-device, no cloud dependency. Latency, privacy, and predictability weren't nice-to-haves. They were the product.
When I joined, the vision was clear but the technical direction wasn't. Key decisions around architecture, cost, and performance would determine whether this shipped as a dependable device or a fragile prototype.
My Role
I owned the end-to-end system architecture and drove execution across hardware, firmware, inference, and product.
Early technical decisions, hands-on hardware iteration, and coordinating across partners to turn an open problem into a production-ready system. The kind of work where product, engineering, and physics all have opinions.
How it works: real-time detection and alerting for people and vehicles.
Key Actions
Edge-First Architecture
Decided early to run inference fully on-device. No cloud dependency. This cut latency, simplified long-term costs, and gave us strong privacy guarantees. Critical for a device operating around children.
Custom Hardware & TPU Integration
Led board-level development and worked directly with Google to integrate the Edge TPU. Optimized models, thermals, and power draw to sustain real-time inference on battery. Every watt mattered.
Productized the Full Stack
Productized the full stack: hardware, firmware, and a supporting web and cloud layer for configuration, updates, and observability. Not a demo. A shippable system.
Results
Delivered a battery-powered device that detects children and vehicles in real time. 60+ FPS on-device inference, no connectivity dependency.
The system behaved predictably under real-world conditions. Reliability, privacy, and performance over theoretical capability. It worked where it needed to, every time.
Key Learnings
Constraints clarify decisions
Designing without the cloud as a fallback forces sharper tradeoffs. Performance, cost, and privacy stop being abstract. They become the constraints you ship against.
Edge AI requires full-stack ownership
Hardware, software, and product decisions were tightly coupled. Progress meant treating the system as a whole, not as separate layers. Silos would've killed it.
Reliability beats complexity
A smaller, predictable system that works every time beats a more capable system that depends on ideal conditions. Especially when kids are involved.