NFO is built to plug into the PLCs you already run — Modbus TCP, OPC UA or EtherNet/IP, read-only — reason over your signals in real time, and hand your operators maintenance calls with the evidence attached. Built and tested on simulated data only; it has not yet run on a plant's real data. The design goal: entirely on your hardware.
Everything above is simulated — a working slice of the console we're building, running on synthetic data in your browser. At a plant, NFO is designed to read your own PLCs and sensors in real time, with the goal of running entirely on your hardware.
One planned install, three role profiles built into the demo (role switching is not tied to login) — book a demo to drive them live.
In the demo: the sensor wall with simulated AI analysis, top risks, risk trend and the action log. Plant composition health up top; every alert one click from its evidence.
In the demo, simulated figures for value delivered, downtime avoided, work orders, energy and quality — per-asset health rolled up to what it costs and what NFO could save.
In the demo: routing and system health — how signals flow to expert packs, how the reasoning mix is weighted, and how deep the system is thinking. Read-only by design.
Built: read-only Modbus TCP, OPC UA and EtherNet/IP connectors, plus replay for commissioning. Tag maps live in plain YAML, with the gateway switched by a single setting. Tested on simulators only.
Built, demo on simulated data: threshold, missing-signal, spike and drift detection; ML anomaly scoring (shadow-only); cross-tag correlation; episode lifecycle — designed to be governed by one fabric.
Built, demo on simulated data: an ongoing condition is one persistent episode with per-episode emit cooldown, and stale episodes auto-close. Severity is meant to be unambiguous: info, warning, critical.
Built, demo on simulated data: upload a manufacturer datasheet and limits are extracted from it. The aim is layout-aware extraction of real spec tables, so recommendations cite rated limits, maintenance intervals and model ids.
Built and released as v2.0.0 and v2.1.0: a per-user Windows installer with no admin prompt, system tray, single-instance lock, and checksum-verified auto-updates. Unsigned, and never installed at a plant.
Built, demo on simulated data: a write-ahead-logged SQLite local store, indexes and a TTL cache. A single batch endpoint is aimed at returning the whole dashboard in under 3 ms — a target, not a measurement.
NFO is being designed to run entirely on-premises, with no cloud tenant, no vendor tunnel, and no outbound AI service touching your plant. The reasoning is planned to run beside your PLCs, not between them and the internet — the aim being to leave malicious-AI attacks on your controllers little to grab onto.
The goal is for all data, models and reasoning to stay inside the building. The initial release is planned to be fully air-gapped — NFO is meant to install and run with no internet connection at all.
NFO's connector code contains no write calls: it listens to Modbus, OPC UA and EtherNet/IP and does not write to the PLC (built; tested on simulators only). The AI is meant to recommend — only your people actuate. Supervised write-back is on the roadmap, opt-in and gated the same way.
Findings and episodes are persisted locally with their evidence (built; demo on simulated data), so you can verify what the AI saw and said.
NFO is built as a desktop product: a per-user Windows installer exists (v2.0.0 and v2.1.0, unsigned, never installed at a plant). The aim: no IT ticket, no server provisioning, no cloud onboarding. Close the window and it is meant to keep observing from the tray.
A pilot is planned to take an afternoon to stand up: one machine, your PLC tags, your datasheets. Keep the findings either way.