SAAM — Seco AI Apps Market

Industrial environments need AI where the machines are — for latency, for resilience when connectivity drops, and above all for data privacy. But deploying, updating, and evaluating AI models on heterogeneous edge hardware is complex, and most industrial users are not AI engineers. SAAM set out to build an ecosystem that makes on-device AI deployable and manageable, including by non-technical users through the no-code Clea AI Studio platform.

Technicians, plant managers, and IT managers dealing with machine faults on the factory floor — plus, at the platform level, the non-technical users that Clea AI Studio’s no-code “AI bricks” are designed to serve. In the medical scenario, clinicians who need image-analysis support without patient data ever leaving the device.

• AI prototyping — offline fault-management assistants for industrial access-control systems, integrating Llama 4, Qwen 2.5, and Qwen 3; experiments with multimodal models (Llama Vision) for analyzing technical schematics and CAD diagrams.
• Model optimization — lightweight LLMs under 8B parameters tuned to run locally on resource-constrained hardware.
• Hardware benchmarking — comparative analysis of edge-computing and federated-learning frameworks; benchmarking models on Qualcomm Snapdragon X Elite and Intel Meteor Lake (Core Ultra) chipsets — tokens per second, hardware-software compatibility — to populate the SAAM Application Hub. Evaluation targets also included the Hailo H10 accelerator and the Edge Impulse platform.
• Workflow and agent integration — n8n workflows automating technical support, RAG systems querying technical manuals in real time during faults, and voice interfaces (speech-to-text / text-to-speech) for industrial and medical machinery.

The work was scenario-driven: each prototype answered a concrete industrial need, then fed measured results back into the SAAM Application Hub.

• Industrial technical support — a system that analyzes error logs, consults the manuals, and suggests targeted fixes to technicians, plant managers, and IT managers.
• Medical imaging (MedGemma) — assisted analysis of chest X-rays to support differential diagnosis (e.g. pneumonia vs. fibrosis), with patient data kept on-device.
• Predictive maintenance — integrating activity logs to improve diagnostics and efficiency of industrial equipment.

• Privacy through locality: sensitive scenarios (medical imaging, industrial diagnostics) run entirely on-device — the deployment model is the privacy model.
• Small models, measured honestly: staying under 8B parameters and publishing per-chipset benchmarks (tokens/second, compatibility) instead of assuming cloud-class models everywhere.
• Offline-first assistance: fault-management prototypes designed to work without connectivity, because faults don’t wait for the network.

The proof of concept was completed: working prototypes across the three scenarios, a populated Application Hub with per-chipset benchmark data on Snapdragon X Elite and Intel Meteor Lake, and validated integration paths with the Clea AI Studio no-code platform — demonstrating that privacy-sensitive industrial diagnostics can run on-device with sub-8B models.

• The full edge lifecycle: taking a model from optimization through containerization to deployment on a specific device is a different discipline from training it.
• A practical toolchain: Ollama, n8n, SQLite, and PyTorch/TensorFlow as the working stack for on-device AI applications.
• Research-industry collaboration: working between university research and SECO’s R&D teams requires translating in both directions — a skill in itself.