AI moved out of the cloud, onto the street.
Architected real-time YOLO on NVIDIA Jetson for a smart-city digital twin.
AI Engineer · architect & lead developer, directing collaborating engineers · Nov 2024 – Apr 2025 · Marrakech
Context
A digital-twin platform for urban services and incident prevention needed to detect and classify urban objects in real time.
Problem
Inference on DigitalOcean cloud was too slow, too costly and not accurate enough; energy leaks were hard to spot; internal documents weren't searchable.
Architecture
- City cameras
- NVIDIA Jetson · YOLO
- Digital twin
- 50+ IoT sensors → anomaly detection
- Archives → OCR → Llama-2 RAG
Key decisions & trade-offs
Edge over cloud
Moved YOLO inference from the cloud to NVIDIA Jetson devices on site.
Trade-offHardware to manage in the field; in return −30% cloud cost, −50% deployment time and +60% detection accuracy.
Automate sensor ingestion
Built an energy pipeline streaming area-level consumption from 50+ IoT sensors at 95% fidelity.
Trade-offA pipeline to maintain, but leaks and anomalies surface automatically.
RAG with OCR for archives
Delivered Llama-2 RAG pipelines with OCR over internal documents for employee self-service.
Trade-offDocument parsing adds complexity; employees answer their own questions.
Outcomes
- +60% detection accuracy vs. the cloud setup
- −30% cloud costs
- −50% deployment time
- 50+ IoT sensors at 95% data fidelity