Arundo Foundation

The heart of Foundation: a typed knowledge graph containing everything about your physical and digital world. Assets and their hierarchies. Sensors and how their data is stored and transformed. Models and computes and what they attach to. Documents, templates — and the agents and apps that read it all through MCP. One graph, one truth — the context layer that makes AI safe and useful in a plant.

Templates are a real inheritance system, not tags: model one vessel or one turbine with typed, validated attributes, then onboard the next ten in one step. And each Domain can run on its own databases — graph, timeseries, and documents — a genuine data-residency answer for regulated operators.

Go deeper: Why there is no Level 2 without a Domain→

Purpose-built for high-velocity, high-volume sensor data from any source — field protocols and plant historians via Arundo Edge, IoT hubs, REST, and file import. Configuration lives in the Domain and is literally in charge: edit a sensor's resolution or freshness rules in the graph, and the datastore's ingestion, validation, and aggregation change with it.

The details are built for trust: queries choose their gap-filling on the fly (interpolate, last-value, zero, or none — no recomputation), every aggregate reports how much was measured versus imputed, and each query is automatically routed to the cheapest storage tier that answers it correctly. Retention, compression, and tiering are policies you set, not projects you run.

ETL that automatically aligns and transforms data into whatever shape your ML models, LLMs, and computations need. Every model input carries its own data contract — period, resolution, aggregation, imputation, unit — so the pipeline is specified by the Domain, per input, with no per-project plumbing.

Alignment is automatic: every input lands on a common time axis, units convert on the way (the model asks for bar, the sensor reports psi), and a resolution policy catches misconfigured inputs before they corrupt a result. Backfill and recalculation are API calls, not database surgery.

Go deeper: Industrial data, analysis-ready→covers capabilities 2 and 3

Dependency-driven parallel execution of thousands of models and computations, autoscaled on queue depth. Five compute types — arithmetic, function, HTTP, external model, and sandboxed Python — run as one graph, with failures cascading only where dependencies demand and optional branches surviving. Designed for exceptional speed and cost-efficiency, because Level 3 means running the whole plant's reasoning around the clock, and that has to be economical.

Every executed tick is a replayable, queryable unit of work: re-run the entire compute graph over any historical period, with a preview before you commit. When an operator asks "why did it say that?", Foundation shows its work.

A complete vector store for your equipment manuals, maintenance logs, procedures, and engineering docs — tied in the Domain to the assets they describe. Retrieval is hybrid, not just similarity: semantic search and full-text search fused, so part numbers and error codes are found as reliably as concepts.

Provenance runs to the exact spot on the page: every answer can cite document, version, page, and the bounding box it came from. Response budgeting is agent-safe by design — a 4,000-page manual never blows up an AI's context.

Go deeper: Document intelligence→

Foundation ships a complete set of MCP servers exposing the Domain, live data, computations, and documents — 24 tools and counting — so your teams work with Claude, ChatGPT, or any MCP-compatible client, directly against the plant. Secured like the platform it is: full OAuth 2.1, every tool call executing as the real signed-in user with their exact permissions.

Analysis is first-class: agents run sandboxed Python (pandas, scikit-learn) in a hardened, egress-locked environment, and interactive Foundation charts render directly inside the AI client. The AI landscape will keep changing; your platform shouldn't have to.

How Arundo stands up your digital twin and first applications at speed: a subject-matter expert describes the plant in plain language, and the kickstarter turns it into a live Foundation domain — databases provisioned, graph imported, history seeded, live data ticking. The same pipeline has materialized everything from a four-sensor demo to a 475-asset, 4,500-sensor plant, deterministically and reproducibly.

Generated applications arrive complete: a shared design system, dashboards bound to live signals, ten languages out of the box — and a built-in AI assistant, global and per-dashboard-card, wired to the Domain through MCP.

A full alerting platform, on from day one: default sensor-health rules are seeded automatically for every new organisation, so the plant starts watching itself the moment data flows. Six rule types feed a real detection state machine; deliveries go to email, Slack, webhooks, and a live in-app inbox with per-user preferences and watch rules.

Reliability is engineered, not hoped for — transactional outbox, idempotency, stalled-delivery recovery — and the framework is open: any system can publish a signal and get the same rule evaluation and delivery machinery. This is the substrate Level 3 recommendations arrive on.