Service
Edge & Federated AI
Inference at the plant or branch, and learning across sites without moving raw data.
Some AI has to run where the data is created.
A camera on a production line can't wait for a round trip to a data centre, a branch may have a weak connection, and many sites aren't allowed to send raw data anywhere. Models then have to run at the edge, be updated across many sites, and keep improving without centralizing the data.
We design and deploy edge inference and, where it fits, federated learning, with the fleet management and monitoring that keep many sites manageable.
You may need this if
- Latency or connectivity rules out sending data to a central server.
- Sites or partners can't share raw data, but want a better shared model.
- Models work in the lab but drift once deployed on site.
- Updating models across many sites is manual and risky.
What we deliver
01
Edge inference
Models optimized for the target hardware (quantization, compilation) and packaged for offline installation.
02
Fleet deployment
Signed model and software updates rolled out in stages, with rollback.
03
Federated learning
Training across sites where only model updates leave each site, with secure aggregation where needed.
04
Monitoring and drift detection
Health, accuracy and data drift per site, visible centrally without collecting raw data.
05
On-site integration
Connections to cameras, PLCs, MES or branch systems through approved interfaces.
06
Handoff
Runbooks for site teams and the central team, plus training.
Reference architecture
A reference design, not a fixed product. Discovery adapts it to your systems and constraints.
Central
- Model registry
- Aggregation server
- Fleet dashboard
Between sites and centre
- Signed updates down
- Model updates and metrics up
- No raw data
Each site
- Edge server or device
- Inference service
- Local training
Sources on site
- Cameras and sensors
- PLC and MES
- Branch systems
Controls
- Code signing
- Staged rollout
- Drift alerts
- Offline mode
How an engagement usually runs
- Discovery2 weeks
Sites, hardware, connectivity and data rules surveyed; an edge, central or federated design chosen.
- Pilot6–8 weeks
One model running at one or a few sites, with updates and monitoring.
- PodMonthly
Rollout to more sites, federated training, and fleet operations.
Constraints we design for
- Limited connectivity
- Sites keep working offline and sync when the link is available.
- Raw data can't leave the site
- Federated or on-site training, so only aggregated model updates move.
- Industrial environments
- Operational network rules, safety reviews and plant change windows.
- Hardware limits
- Models sized for the devices on site, not the lab.
Typical stack
- ONNX Runtime
- TensorRT
- OpenVINO
- Flower
- K3s
- MQTT
- Python
- C++
We use the tools you already run where they fit. Names are for reference only.
Tell us about your systems
An engineer reads every message and replies within one business day.