SAMBA — Small Language Models Enabling Industrial IoT-Based Agentic AI. On-premise agents that read plant data, investigate events and propose responses. People with authority decide.
Agents read data, investigate events and propose responses. Nothing changes in the plant without the approval of the people responsible.
The challenge, the approach, the pilots and the consortium — explained in one short presentation.
Video v1.0 · 7:44 · English. Narration produced with an AI-generated voice. Open the video in a new tab.
Plants are full of data, but turning an event into a sound, approved decision is still manual work across MES, ERP, WMS, SCADA and AGV screens.
Alarms are not linked to orders, materials, lots and schedules. Root-cause and impact analysis is done by hand.
No industrial profile defines how agents may use MES, ERP, WMS or AGV functions safely. Adapters are rewritten at every site.
Data sovereignty, segmented OT networks and continuity require on-site models — and reliable tool use by small models is unproven.
Recommendations are not tied to evidence, not checked against plant rules and not logged in a coherent audit trail.
Disruptions spanning areas, AGV fleets, plants or companies are handled by meetings and email. Centralising data is not an option.
SAMBA adds supervised agents on top of existing systems — without replacing PLCs, SCADA, MES, ERP or WMS. Plant data stays on site.
None of them is useful on its own. Together they take an agent from a plant event to a checked, approved and logged action.
Each tool declares safety level, roles, approval rules, returned evidence and mandatory logging, mapped to OPC UA nodes or AAS submodels. Open specification proposed to the MCP community.
Hybrid set-up with solvers and detectors; adaptation by task generation, distillation, PEFT and quantisation; an open industrial agent benchmark.
Area agents make structured proposals; a coordinator detects conflicts over material, AGVs or capacity and resolves them with solvers and priority rules — across sites too.
Deterministic checks outside the model decide what is shown. Two levels of explanation (operator and auditor), approval routing and full decision logs aligned with the AI Act.
Each pilot starts with read-only or recommend-only agents and moves to supervised actions only where the plant allows it.
A late part, a degraded press or an urgent order is re-planned by hand across MES, ERP, warehouse and AGV screens. Agents re-sequence assembly and propose AGV missions and maintenance slots.
On an order change, stoppage, late truck or near-expiry lot, agents call solvers and propose picking, sequencing or dock slots — checked against FEFO and lot rules.
Continuous reconciliation of silo stock against electronic warehouse receipts, with explained alerts when a discharge doesn't match level change and weighings.
Agents separate measurement faults from real events and propose fixes, transfers, returns or service tasks for drums, racks and containers.
When a CNC stage slows down, agents identify the bottleneck, estimate the orders affected and propose re-sequencing, inspection or shifting capacity between factories.
Agents gather linked evidence on thread-rolling cracks, ask for what is missing and propose a cause hypothesis — checked by a separate verifier before the quality engineer decides.
Sites that cannot pool data turn local deviations into structured summaries; agents on other nodes check their exposure and a coordinator combines the proposals.
Baselines are taken in the first months of each pilot and compared at the end against the same fixed reference models.
These are project targets, not achieved results. * Indicative target, to be fixed after the baseline.
An open profile specification plus commercial modules — on-premise licences, subscriptions and integration services from the partners.
| Output | What it is | Target TRL |
|---|---|---|
| O1 Industrial MCP profile & reference servers | How industrial functions are described for agents — safety level, permissions, approval rules, mapping to OPC UA and AAS — with servers for MES/MOM, ERP, WMS, SCADA, AGV and tracking. Profile published openly. | 6 |
| O2 Edge SLM toolkit & industrial agent benchmark | Methods to adapt open SLMs in a hybrid scheme with solvers and detectors; benchmark of tool-call accuracy, consistency, response time and energy, with a published protocol. | 5–6 |
| O3 Supervised multi-agent coordination framework | Area and site agents with a coordinator; handling of competing proposals; controlled exchange of context summaries without sharing raw data. | 5 |
| O4 Validation, explanation & audit module | Rule-based checks before a proposal is shown; explanations linked to measurements, orders and procedures; approval flow; decision log for AI Act documentation. | 6 |
| O5 SAMBA Agentic AI Platform | On-premise integration of O1–O4 with OPC UA and MQTT connectors, deployable on edge servers without retiring existing systems. Open-source core with commercial partner extensions. | 6 |
Industrial pilot owners, technology SMEs, universities and research centres. Coordinated by Alpata (Türkiye).
Ullswater Investments owns the Spanish pilot and leads TI1, bringing on-premise SLM agents that act on MES, ERP, warehouse and AGVs through MCP tools — always under human supervision.
Read. Agents read the order queue, stock, AGV positions and shift, and investigate disruptions in cabin and door assembly and warehouse logistics.
Recommend. A solver re-sequences assembly; agents propose the sequence, AGV missions and a maintenance slot, with evidence attached.
Execute with approval. Approved actions are re-checked before execution and logged end to end.
Manufacturers, logistics operators and technology providers in Spain can contact the Spanish coordination team to learn more about the project and its pilots.