- Client
- Product start-up (under NDA)
- Platform
- NPU-based Linux single-board computer
- Deadline
- Customer demo in 36 hours
- Delivered
- Camera → NPU detection → app output
The problem
The client had a customer demo booked. The product needed a camera, real-time object detection on an NPU, and results sent to their application. Two days before the demo, their developer backed out. What remained was a board, a trained model and an empty repository.
Give me the board, the model and the demo script. We will plan the next 36 hours around what the customer needs to see.
How I planned 36 hours
With no time for exploration, I cut scope to what the demo needed and did the riskiest work first.
Hours 0–4: prove the NPU path
Converted the model with the vendor toolkit, quantised it with sample frames and confirmed it ran on the NPU with correct outputs. Everything else depended on this.
Hours 4–12: camera pipeline
Brought up the camera with hardware-accelerated capture and resize, so frames reached the NPU without wasting CPU.
Hours 12–24: application
Built the inference loop, post-processing, tracking and a lightweight interface that sends results to the client's app.
Hours 24–32: harden
Auto-start on boot, recovery if the camera disconnects, logging, and a long soak test.
Hours 32–36: rehearse
Ran the full demo script with the client twice and fixed the last rough edges.
What the client got
Results
- A working edge AI demo, delivered on time.
- Real-time detection at about 25 frames per second, fully on the device.
- Clean, documented code they could continue to build on after the demo.
- A follow-up plan to turn the demo into a production build.
Embedded systems consultant at TechProgeny, Bengaluru. 7+ years and 29+ shipped products across firmware, embedded Linux, edge AI, apps and IoT servers.