# The Four Layer Stack and What Is Commodity ## The architecture is sensible, but anyone can get most of it Foreman has four layers. The **infrastructure** layer has on-premise, private or hybrid cloud, and GPU inference. The **data** layer has documents, databases, ERP and MES connectors, and a vector index. The **agent** layer has a planner, task decomposition, tool calling, a human approval step and guardrails. The **app** layer has a chat interface, a low-code builder and tool endpoints. The layers have the correct shape for an on-premise-first system, because the company can replace the model behind the agent layer. The company does not change the agent logic for this. > [!success] What is special > The company has classical manufacturing machine learning, which is connected to language-model agents. The machine learning includes root-cause analysis, virtual metrology, parameter optimization and explainability. Most horizontal agent platforms do not have it. It agrees with the applied background of the founder. > [!failure] What is not special > This part has agent orchestration with task decomposition. It has hybrid (keyword plus semantic) retrieval, text-to-SQL and guardrails. It has a low-code workflow builder and containerized model serving. A competent team can assemble this layer from public components in **months** (estimate of the reviewer). The cost is in the low hundreds of thousands of dollars. ## Where defensibility can be | Candidate | Status | | --- | --- | | Curated manufacturing knowledge bases and schemas | The value is in them, but partners supply them, see [[Borrowed Data Is Not a Moat]] | | Classical manufacturing ML modules | The strongest asset. The company gave no model inventory and no metrics | | Knowledge-capture ("memory") engine | The company mentions it. The company does not show it and gives no information about it | | Domain packs for each vertical | The data show them in one vertical only, see [[Domain Packs and the Vertical Playbook]] | > [!warning] The architecture diagram > The company description names some integrations: shop-floor control systems, historians and IoT protocols. The architecture diagram does not show them. The diagram also does not show vision models. It is possible that vision models are necessary for the largest contracts. See [[Claims Versus Evidence]]. ## Why it matters The deck shows the platform layer at the front. The defensible value is in the domain layer and the ML layer. These layers have the minimum documentation. The diligence team must spend its time on these layers. The frame in [[Technical Moat Assessment Framework]] gives the method. ## Related - [[Foreman MOC]] - [[Open-Source Dependencies and Licence Exposure]] - [[AI Agents Stack]] - [[Sensor Data and Historian Systems]]