Service — 04 / AI

Artificial Intelligence
applied to
real industry.

We build AI applied to industry: custom MCP servers that give LLMs scoped access to third-party industrial systems, AI-native products that solve specific process problems, disciplined prompt engineering for technical assistants and code analyzers. Computer vision and predictive maintenance when the problem requires it.

— 01 / Approach

Three principles on how we apply AI.

AI solves real industrial problems, it doesn't fill slides. Integrated with existing systems, designed with rigour, validated on the client's process KPI.

— 01

MCP-first: AI that acts

We build custom MCP (Model Context Protocol) servers giving LLMs scoped, controlled access to ERP, PLM, SCADA, industrial databases, PLC code. AI doesn't just read PDFs — it queries the real systems and, where it makes sense, runs constrained actions.

— 02

Disciplined prompt engineering

Access to an LLM is not enough. We design prompts, retrieval, guardrails and tool calling per use case, testing them on real client scenarios. We have specific in-house expertise — prompt engineering is a craft.

— 03

AI as a tool, not as the point

Every AI application solves a concrete problem (understanding a fault, predicting a stop, tracing a part, supporting a maintenance engineer). Our solutions don't force AI in at any cost — we integrate it when it's genuinely needed. Value is measured in time saved and errors avoided.

— Product / ENTERPRISE AI PLATFORM

PaLoMa AI

Build AI agents on your systems. In days, not months.

Standalone platform for chatbots and AI agents over your data and processes. Multi-LLM, multimodal RAG, custom MCP on demand, enterprise governance, no-code admin — out-of-the-box, no vendor lock-in.

  • Custom MCP on demand over your systems
  • Multimodal RAG and enterprise governance built-in
  • Already in production at an automotive company
Discover PaLoMa
PaLoMa AI
— 02 / Tech stack

The technologies behind our AI solutions.

Pragmatic, vendor-neutral where it makes sense. We pick the model and architecture that fit the use case — there is no universal answer.

/ LLM

Foundation models

The engines. Picked by use case, cost, data constraints.

Azure OpenAI (GPT-4o)Anthropic ClaudeMistralLlama (self-hosted)
/ MCP

AI ↔ systems integration

Custom MCP servers, function calling, agents that operate on client systems.

MCP SDKLangChainLlamaIndexFunction callingAnthropic SDK
/ RAG

Retrieval & vector DB

Company-knowledge memory: technical docs, manuals, regulations.

pgvectorPineconeWeaviateOpenAI embeddingsCohere
/ Vision · ML

Computer Vision & classical ML

When the problem does not need an LLM: detection, classification, time series.

PyTorchYOLO (Ultralytics)OpenCVXGBoostscikit-learn
/ Deploy

MLOps & inference

Versioned models, reproducible deploys, on-edge or cloud inference.

MLflowNVIDIA TritonNVIDIA JetsonDockerAzure ML
— 03 / Real cases
— 01 / Tracking Code VISION · TRACEABILITY

Product traceability via computer vision

Industrial vision system that recognises product codes — printed or labelled — on moving parts in real time across production lines, buffers, and warehouse entries/exits. Every transition is logged automatically. FITEC internal POC: Tracking Code.

YOLOOpenCVPyTorchTensorRTNVIDIA Jetson
— 02 / Predictive TIME SERIES · ML

Line-failure prediction

Machine learning models trained on sensor streams (vibration, temperature, current, pressure) of a production line to predict imminent line blockages and compromised parts before they occur. Streaming anomaly detection + classification of the failure type, with useful lead time for the maintenance team.

XGBoostPyTorch (LSTM)Apache SparkProphetMLflow
— 03 / PLC Assistant LLM · MCP · RAG

AI assistant for machine stoppages

During a line stoppage the assistant analyses in real time the plant's PLC code, cross-referencing it with the line's technical documentation. The maintenance engineer asks in natural language 'why did it stop?' and the AI replies citing the specific PLC code lines involved, the critical I/O and the documented procedures. RAG over docs.

Azure OpenAIAnthropic ClaudeMCP SDKpgvectorTIA / TwinCAT bridge
tracking · live CAM CAM.01 042 118 239 476 591 SKU-A742-27 SKU-B118-04 SKU-C239-92 — SCAN LOG 23:04:18 · A742-27 23:04:14 · B118-04 23:04:11 · A742-26 23:04:07 · C239-92 23:04:03 · A742-25 23:03:59 · D481-77 23:03:55 · B118-03 23:03:51 · A742-24 23:03:48 · C239-91 23:03:44 · D481-76 DETECTED 1.247 RATE 180 / min ACCURACY 99.4%
sensors · live stream ALERT — SENSOR STREAM · LAST 60 MIN ANOMALY DETECTED vib.X vib.Y temp curr — RISK SCORE 84% HIGH — PREDICTED FAILURE Bearing failure · spindle 02 ETA · 4.2h CONFIDENCE · 92% ACTION · maintenance.alert.sent
plc-analyzer · line 3 LINE STOP OPERATOR · 14:23:50 linea 3 in fermo, perché? AI · 14:23:53 Allarme E202 — timeout valvola V14 nella sequenza di carico (blocco RB-23 ). L'output Q3.4 non rileva chiusura entro 5s. [1] plc/line3/ladder.txt:144 [2] doc/lin3/manutenzione.md §3.2 [3] schema/v14_circuit.pdf checking i/o state — PLC CODE · ladder.txt 142 IF V14_OUT THEN 143 START_TIMER(T14) 144 WAIT V14_CLOSED, 5s ← TIMEOUT 145 ELSE FAULT(E202) 146 END_IF 147 148 // sequence: load_RB23 149 // see DOC-LIN3-V14 — MCP TOOL CALLS read_plc_state(line_3) 87ms search_docs("E202", "V14") 142ms read_ladder(line_3, 140..150) 34ms cross_reference(V14, RB-23) 61ms summarize(context) — DIAGNOSIS · ready Timeout valvola V14 · sostituire o pulire elettrovalvola CONNECTED · 3 MCP servers TOOL CALLS · 5 TIME · 2.4s CONFIDENCE · 94%
— 04 / How we operate

From problem to AI in production.

Five stages, short iterations. We validate against process value at every step — we don't build a model unless we're confident it will move the right KPI.

— 01
Discovery & problem framing
We check whether it's actually an AI problem. For many cases a good heuristic solves it better and with less complexity. When AI is the right fit, we define measurable process KPIs and a baseline.
— 02
Data & context audit
For ML/Vision cases: available data, quality, labelling. For LLM/MCP cases: catalogue of knowledge (technical docs, code, accessible systems), schemas of the target systems to integrate.
— 03
Design & prompt engineering
Solution architecture: model choice, prompt and tools design, retrieval strategy, guardrails. For vision/ML cases: architecture choice and baseline. Fast iteration with real-user feedback.
— 04
Integration & deploy
MCP servers connected to the systems (ERP, PLM, SCADA, PLC bridge). On-edge inference for vision cases, cloud API for LLM flows. End-to-end tests with real client scenarios.
— 05
Monitor, eval, evolve
Continuous output-quality eval (for LLM: prompt regression tests; for ML: drift detection). Versioning of prompts + models, A/B testing on new releases. The live system evolves with the client.
— 05 / FAQ

Frequently asked.

— 01 What is MCP and why do you use it?
MCP (Model Context Protocol) is an open standard that lets LLMs access external systems in a structured way — read data, call functions, run constrained actions. For us it's the clean way to integrate AI into the client's industrial systems (ERP, PLM, SCADA, PLC) without inventing one-off custom integrations.
— 02 Are you tied to a specific LLM?
No, we're not. We design the architecture so the underlying model can be swapped, run comparative tests on real client scenarios, and evaluate compliance, cost and data constraints (e.g. on-prem, data sovereignty) to pick the best LLM for the task at hand.
— 03 Is prompt engineering really a craft?
Yes, and it's where most AI projects fail silently. A poorly designed prompt with a good model gives apparently acceptable demo results but breaks in production. We invest in prompt design, systematic testing, guardrails and automated evals — it's not 'write better instructions', it's engineering.
— 03 / Contact

Got an industrial process to digitize?

Tell us the problem. We'll design the right solution together, from proof-of-concept to production.

Write us