GUS.ai
Industrial AI.
An agentic co-pilot that reads your SOPs, watches your sensors, and traces your PLC. Every answer comes back with cited evidence.
What is GUS.ai
Three knowledge sources.
One conversation.
Most plants split their knowledge across three silos: documents nobody re-reads, dashboards nobody watches, and PLCs nobody opens. GUS.ai pulls all three into a single agentic loop, with cited evidence on every answer.
Layer 1
Unstructured documents
SOPs, P&IDs, equipment manuals, recipe specs, parts lists, commissioning reports. Embedded into a vector store with chunk-level provenance so every answer cites a source, page, and section.
Layer 2
Live process telemetry
Real-time sensor and equipment data streamed from PLCs and instruments through HiveMQ and TimescaleDB. Current values, trends, alarms, batch state.
Layer 3
PLC ladder logic
Allen-Bradley ControlLogix programs parsed from L5X — UDTs, controller tags, alarm definitions, and routine references — indexed and answerable in plain English.
Chapter 01
What
GUS.ai does.
Capabilities, all running on the same agentic loop.
Capabilities
What GUS.ai can do.
14 agentic tools, 60+ API endpoints, 5 communication channels, streaming chain-of-thought — built to answer the questions plant engineers actually ask.
Document Intelligence
Semantic search across SOPs, P&IDs, equipment manuals, recipe specs, parts lists, and commissioning reports. Every answer cites the exact source, page, and section.
~700 chunks · 8 ingest pipelines
Live Telemetry
Real-time sensor data from your process. Current values, time-series trends, and threshold monitoring across every instrumented point in your plant.
Curated catalog · live tag hierarchy from Chronicler
Batch Analytics
Step timing analysis, batch history comparison, spec-vs-actual deviation detection. Automatic batch boundary detection without an MES — discovered from raw tag transitions.
Duration · temp milestones · cross-batch trends
PLC Intelligence
Parsed L5X programs — UDTs, controller tags, and alarm definitions indexed as RAG chunks. Ask about any rung, any tag, any alarm in natural language.
Allen-Bradley ControlLogix · L5X parser
Process Intelligence
Anomaly detection using MAD-based z-scores with coincident spike grouping. Multi-tag Pearson correlation analysis. Impact propagation, root cause analysis, and single-point-of-failure detection across the plant dependency graph.
Anomaly · correlation · causal reasoning
Multi-Channel Access
Ask GUS.ai from the web app, by email, by text message, by phone call, or by Telegram. Same agentic pipeline on every channel — wherever you are on the plant floor or off it.
Web · Email (AgentMail) · SMS · Voice · Telegram
Proactive Alerts
Automatic notifications on batch start/end and CIP transitions. Heartbeat monitor sends alerts on alarms, step changes, and abnormal conditions. System-status monitor with incident and recovery emails.
Batch alerts · heartbeat · status monitor
Audit & Compliance
Tamper-evident audit trail of every query, tool call, and response — SHA-256 hash-chained, append-only. Daily digest emails. Golden-question management for QC. Full CSV export.
Hash-chain · digest · CSV export · Clerk JWT
Visualization
See your plant,
your code, your data.
Eight purpose-built visualizations — each tied back to the same agentic loop. Click anything, ask anything, get back cited answers.
Code graph
Navigate ~900 backend code entities in 3D. Click any node for an AI-generated purpose description; ask follow-ups in the side panel.
Knowledge graph
Auto-built from RAG chunks via LLM extraction. See how SOPs, manuals, and commissioning reports relate.
Docs graph
Internal docs as a navigable network — from CLAUDE.md to ARCHITECTURE.md to per-module READMEs. Chat with any node.
Dependency graph
Static plant-process model — currently 128 nodes and 155 edges. Backbone of impact, RCA, and SPOF reasoning.
Data flow graph
Radial ADX hub: instruments and PLC fan inward through edge gateways and HiveMQ to the historian and GUS. Live MQTT pulses cascade in real time.
MQTT explorer
Live topic-tree subscriber over WSS to HiveMQ Cloud. Diff, flash-on-update, history, resizable panels. Read-only.
Digital twin
Real-time vat overlay with 14 extended telemetry tags. Watch a batch run alongside the AI answers about it.
Causal reasoning panel
19 pre-built questions across four categories — impact analysis, root cause, SPOF detection, process health — plus a free-form input.
Chapter 02
Built on
serious AI.
Right model for the right job, with safeguards on every channel.
AI architecture
Right model for
the right job.
GUS.ai uses five specialized models from two providers — each chosen for a specific task, balancing reasoning depth, latency, and cost. An OpenAI GPT model stands by as a Claude failover.
Powered by Claude
Claude Opus
Highest reasoningDeep reasoning
Powers the interactive code, knowledge, docs, and dependency graph conversations — structural analysis and architectural reasoning across the entire codebase and plant model.
Claude Sonnet
14-tool orchestrationCore agent brain
Primary orchestration. Manages a 14-tool agentic loop with extended thinking — deciding which documents to search, which tags to query, and how to synthesize evidence into cited answers.
Claude Haiku
Low latencyReal-time descriptions
Generates instant purpose descriptions for graph nodes on click. Optimized for sub-second latency with server-side response caching.
Automatic OpenAI failover for 99.9% uptime
Powered by OpenAI
Whisper
Speech recognitionSpeech-to-text
Transcribes spoken questions into text for the agent pipeline. Operators can ask GUS.ai hands-free from the plant floor via voice, phone, or kiosk microphone.
TTS-1
Voice outputText-to-speech
Reads answers aloud with natural voice synthesis. Enables hands-free operation during production runs and equipment inspections.
Security & trust
Built for the
plant floor, not a sandbox.
An AI co-pilot for industrial work has to earn the trust of a control engineer before it earns the trust of a CEO. These are the safeguards we ship by default.
Evidence-mandatory architecture
Every answer cites a source. Every tool call is logged. Every change is verifiable from a tamper-evident chain.
Read-only by default
GUS.ai never writes to your PLC. The agent has no tool that can modify a setpoint, force a tag, or change a recipe. Plant safety is enforced at the tool layer, not by prompt.
Tamper-evident audit trail
Every query, tool call, and response is appended to a SHA-256 hash-chain — break a record and the whole chain fails to verify. Daily digest emails. CSV export.
Boot-time integrity manifest
Config, tool schemas, safety regexes, and the sandbox boundary are SHA-256 manifested at startup. Drift triggers a hard refusal to serve traffic.
Emergency kill-switch
Admin-gated endpoint flips an instant 503 across all channels — web, email, SMS, voice, Telegram. One header to halt the whole co-pilot.
Per-channel tool allowlist
SMS and voice see a smaller subset of tools than web. Email is scrubbed for prompt-injection attempts at the gateway. Defense in depth.
Post-quantum-ready transport
TLS terminates at Fastly's edge, which already negotiates X25519MLKEM768 hybrid PQ. Recipe IP stays sealed against harvest-now-decrypt-later threats.
Chapter 03
Where it
goes from here.
Today's deployment is the beginning, not the destination.
Roadmap
What's coming next.
GUS.ai is actively evolving. Each milestone is informed by real plant-floor feedback — not a market plan.
Research
Predictive intelligence
Edge-side spiking neural networks (Akida, Loihi) for bearing wear, CIP fouling, and batch-phase perception. Retrospective-data prototypes scoped before any hardware purchase.
Designed
Agentic memory
Persistent operator-scoped memory layer with safety controls and an MCP interface. Lets GUS recall plant-specific context across sessions without crossing tenants.
Planned
Specialist agents
Per-domain sub-agents (CIP, batch QA, alarm triage) that share the GUS knowledge fabric but specialize their reasoning loops. Coordinated via a parent orchestrator.
Planned
Multi-vendor PLC
Today: Allen-Bradley ControlLogix via L5X. Next: Siemens TIA Portal, Beckhoff TwinCAT, Schneider EcoStruxure — same parser pattern, same agent surface.
Planned
Mobile PWA + air-gap
Operator-friendly mobile shell for plant-floor use, plus a fully air-gapped on-prem deployment profile for plants with no outbound internet.
Designed
GraphRAG expansion
Move from chunk-level retrieval to graph-aware retrieval — exploit the relationships GUS has already extracted between documents, equipment, and process steps.
About APT
Process engineers
who build AI.
Advanced Process Technologies has spent 25 years engineering dairy and food-processing systems for some of the largest plants in North America. GUS.ai is what happens when that domain knowledge meets agentic AI.
We don’t hire AI generalists and hope they figure out cheesemaking. We started with a master cheesemaker and wired AI around what he already knows.
GUS.ai is named after Mark Gustafson, APT’s master cheesemaker and the domain-expert inspiration for the product.
Quick facts
- Founded
- 2000
- Ownership
- Employee-owned (ESOP)
- HQ
- Cokato, Minnesota
- Patents
- 3 U.S. patents on the ACV
- Reference deployment
- Advanced Cheese Vat
- Named for
- Mark Gustafson, master cheesemaker
See it on a real vat
Bring GUS.ai to your plant.
A 30-minute walkthrough of the live Advanced Cheese Vat deployment. We’ll show you the agent in action against real plant data, not a sandbox demo.
Greg McMillan · [email protected] · APT, Cokato MN