MLOps · Edge AI Fleet
MLOps platform for a 600-device edge AI fleet
End-to-end MLOps platform managing 600 Jetson inspection cameras across 14 sites — median model deploy went from 9 days to 38 minutes, with automatic drift-triggered rollback.
Edge AI · On-device inspection
Zero-cloud defect detection camera on ESP32-S3, case study.
For an electronics-assembly client, Yantrix designed a custom inspection camera that flags solder and placement defects on-device. The whole pipeline — sensor, model, firmware, enclosure — fits in a fanless 40 x 40 x 25 mm module.
By Yantrix Engineering · Edge AI Studio2 min readElectronics manufacturing
Overview
Why this study matters
A production conveyor inspection camera running a quantized INT8 CNN entirely on an ESP32-S3 — 18 FPS at 0.4 W, no cloud, 6× lower capex per station.
Client: An Indian electronics manufacturing services (EMS) firm
Project Type: Edge AI + Custom Hardware
Industry: Electronics manufacturing
Service Used: Edge AI + PCB Design + Embedded Firmware
Results in numbers
What the engagement actually shipped.
- 18 FPS
- Continuous on-device inference
- 0.4 W
- Steady-state power draw
- 96.8%
- Validation accuracy
- 6×
- Lower capex per station
Objectives
What the project needed to achieve
- Detect defined defect classes on placed components with production-grade accuracy
- Run the model entirely on microcontroller-class hardware
- Eliminate cloud and on-prem PC dependencies
- Fit into a compact, fanless enclosure mountable above the conveyor
Challenge
Engineering constraint
The client had a line-side PC-based vision rig that was expensive to scale across many stations and introduced a cloud dependency the plant IT policy didn’t allow. They wanted a self-contained inspection module per station — low power, low cost, zero external dependencies.
Approach
How Yantrix approached the work
- 01Captured a labelled dataset of nominal and defective assemblies across lighting variants at the client plant.
- 02Designed a compact CNN sized for ESP32-S3 constraints and trained it with aggressive INT8 quantization-aware training.
- 03Designed a custom 4-layer PCB combining ESP32-S3-WROOM, an OV5640 camera module, a local ring light, and MQTT telemetry over Wi-Fi.
- 04Delivered firmware with a ring buffer for failure-case capture so the client can periodically retrain from real edge cases.
Outcomes
What improved by the end
- 18 FPS continuous inference at approximately 0.4 W steady-state
- 96.8% accuracy on the validation set; on-line operator override provided for ambiguous cases
- No PC, no cloud — fully autonomous per-station operation
- Unit cost reduced to a fraction of the previous PC-based rig
- Failure-case ring buffer enables continuous dataset growth
Deliverables
What the client receives
- Custom PCB design files and fabrication package
- Quantized INT8 model and training pipeline
- FreeRTOS firmware with OTA update path
- Enclosure CAD with thermal and mounting study
- Commissioning report and retraining playbook
Tools used
Stack and tooling
- ESP32-S3-WROOM-1 (8 MB PSRAM)
- OV5640 camera module
- TFLite Micro with ESP-DL acceleration
- PyTorch (quantization-aware training)
- KiCad for PCB design
- FreeRTOS firmware
- MQTT for telemetry
Impact
Business-level effect
- Inspection coverage extended from selected stations to the entire line
- Capex per station reduced by roughly 6× vs. the PC-based alternative
- Plant IT policy satisfied — no external network traffic required
Conclusion
The project shows that thoughtfully quantized models plus hardware co-design can put real-time ML into truly constrained devices — not just on single-board computers.
Next step
Need an inspection system that scales across dozens of stations without a PC fleet? We design the device, the model, and the firmware as one thing.
Tagged
- ESP32-S3
- TFLite Micro
- Edge AI
- Quantization
- PCB Design
- Defect Detection
Frequently asked questions
Answers from the engagement itself.
Can you really run a useful vision model on an ESP32-S3?
Yes — for narrow, well-scoped problems. A 320×320 INT8 CNN with under ~1.5M parameters runs at 15–20 FPS on the ESP32-S3 with ESP-DL acceleration. Defect detection, presence/absence checks, and SKU classification all fit. Object detection across many classes does not.
How does an ESP32-S3 inspection module compare on cost to a PC-based rig?
BOM landed around ₹6,000–12,000 per station versus ₹75,000–1,50,000 for a PC + industrial camera setup. Power per station drops from ~80 W to under 1 W, which compounds over hundreds of stations across a plant.
Related case studies
Adjacent proof you can read next.
Applied AI · Vision-guided robotics
Vision-guided bin picking at 80 ms end-to-end
How a YOLOv11-Seg + 3D-pose stack on a Jetson Orin Nano replaced fixed-pose jigs in a 6-DOF robotic cell — sub-80 ms latency, 99.2% accuracy, 40% throughput gain.
Continue exploring
Related blogs, services, and capability pages
Cross-links help readers move from proof into capability and educational content, and they reinforce the crawl path between commercial pages and reference content.
Let's build
Have a machine to build? Let's scope it together.
Tell us about your project. We'll respond within 1-2 business days with a preliminary scope and timeline — no boilerplate, no up-sell.