Product shape
Three layers, sold as one "AI OS". Read left to right: data comes in, gets modelled, an LLM agent reasons over it with tools, and a human decides.
Model names are read from the product UI in Juna's demo videos. The enterprise page says customers can choose any LLM.
Foundation — the company model
Juna says its engineers plus roughly four hours of customer input produce a structured map of every entity, data source and flow. Every screen below is paused from the page's product videos.
Connector catalog
- Grouped by layer: real-time process & control, time-series historians, MES & production ops, ERP.
- Named tiles: OPC UA, MQTT, Modbus, AVEVA PI, Siemens Opcenter, Tulip, SAP, Oracle JD Edwards, Microsoft Dynamics 365.
- Tulip appears as a data source — frontline apps feed Juna's analysis.
Self-serve OT connection
- OPC UA form: endpoint URL, security mode (None / Sign / Sign & Encrypt), authentication, root node for sensor discovery.
- The dialog says it stages the endpoint for a later live connector, so the live link may still be engineer-assisted.
Process flow graph
- Machine sections as nodes; material flows (paper stock, starch, vacuum) as coloured edges.
- Filter by material to isolate one flow through the line.
- Demo plant is a paper machine (Stock prep → Refiner → Wire, Press, Dryer sections).
Machine panel
Click a node: outgoing connections, 216 sensors, machine ID, type, location, linked documents. A "Chat" tab is docked on the right edge.
Sensor panel
Current value, range picker (1D to All), min / avg / max, sensor ID and type (here an SO₂ concentration tag).
Knowledge base — library
- Upload manuals, BOMs, SOPs; each file gets an AI summary, status and uploader.
- Juna claims extraction takes under a minute per document, working from PDFs, photos of handwritten notes, voice recordings and group chats.
Original view
A BOM CSV maps machine → functional group → component → historian tag (e.g. W2_UNW_TEN_PV). This is how documents bind to live sensor data.
"Juna Document" view
Summary, topics, and typed entities (equipment, organisation, location, standard, person), plus links to other files: same entity, similar topic. Editable.
Document graph
- Relationship types: part of, same entity, shared equipment, similar topic.
- Topic clusters group documents by shared subjects.
Memory of plant language
- User says "heavy liner" means the 175 g/m² kraft grade, run code HL175; the agent stores it and reuses it.
- Status chips: "Thinking completed", "Juna Agent is programming". Context meter shows 2% used at 21,592 tokens.
- Model picker shows Claude Sonnet 4.6; voice input mic in the composer.
Chat — natural-language analysis on live data
Five advertised use cases: data science, best operating conditions, root cause analysis, reports, dashboards. Answers are charts and quantified recommendations, not just text.
Factory Chat
- Single prompt box with model picker (Claude Opus 4.6), tool grid, starter questions.
- Example ask: which reactor settings gave the best results for a product last quarter.
Data science charts
- Interactive charts with hover tooltips: 7-day power by line at 15-min resolution, polar energy share, efficiency scatter, weekly heatmap.
- Chart titles are phrased as the business question being answered.
Root cause — evidence
Heater draws more power while temperature drops: a fouling signature over 21 days, with setpoint and quality-floor bands. "2 Actions completed", JSON view, Save to Dashboard.
Root cause — money and next step
Converts drift into lost tonnes and €/month, compares "do nothing" with "descale this weekend" (payback in days), then proposes a standing early-warning monitor.
Best operating conditions
Centerlining widget: value vs reference mean and ±2σ bounds across product grades and sheet breaks. Buttons: Send for Review, Disable, Set Reference Bounds. Chat sits beside it.
Diagnosis → suggested actions
Agent concludes a stale reference, not a process fault, and points the user to the exact UI control, with "Send for Review" when team approval is required.
Tools as toggleable "apps"
- 19 apps per chat: Gantt, Heatmap, Knowledge Base, Memory, PDF Reports, PI Data (AVEVA PI), PPT Reports, Production Data, Sensor Data and more.
- Memory is described as "Cockpit memory records" — likely an internal product name.
Reports on demand
Weekly Plant Performance Report as a paged PDF: alarm, PM and MTBF cards, equipment issue table with severity and action. PPTX decks with editable source too.
Live dashboards
Charts saved from chat become tiles; dashboards are tabbed ("Asphalt Works", "Key KPIs 2026") with copy, edit, delete per tile.
Applications — agents that work without being asked
The page claims a library of 250 pre-built workflows. The hero video uses an NH3 refrigeration plant.
Alert inbox
- Alert: freezer tunnel idle while NH3 Circuit B stays at production setpoint; excess cooling for 3 h 47 min.
- Analytics: actual vs expected kWe, compressor load by circuit.
- Recommendation: raise evaporation setpoint −32 °C → −28 °C, est. 18–22 kWe saved. Forward or Override.
- Report due card: auto-generated shift energy report, Review or Send to Team.
Human in the loop
Forward drafts an editable message to the process engineer. The AI never changes the setpoint itself.
Value shown to the user
An inbox-zero streak with hours saved this week, alerts resolved, reports sent on time. Outcome of the action itself is not shown.
Background agents
- Cards with cadence chips: every 15 min (trip investigator), daily 05:30 / 06:00 / 14:00, weekly Monday 07:00, hourly gas efficiency.
- Day-ahead electricity agent pulls ENTSO-E prices against the production schedule.
- Mixed German/English descriptions — authored per customer. "New Agent" button for self-build.
Delivered where people work
Microsoft Teams toast from "Juna agent": a quality parameter trending out of spec on line 5.
Agent email
"Quality Drift Agent" email: SO₂ warning, last 3 lab samples vs spec, current reactor setpoints, likely causes. Carbon-black vocabulary (N330, BET, OAN).
Company vs personal
New dashboard: Company (shared with everyone) or Personal (only you).
Save chart to dashboard
Pick a company or personal dashboard as the destination. Chats can also be shared with a colleague, a team or the company.
Analysis
Fact = from Juna's site, product UI or press. View = our assessment.
Platform Fact
- Three layers: Foundation (company model, knowledge base, memory), Chat, Applications. Launched as the "Agentic Factory OS", positioned as Juna 2.0 (Mar 2026).
- Named agents: Production Scheduling, Process Engineering, Resource Intelligence, Background Agents, Agent Builder on an Agent SDK.
- Private servers on Azure hosted in Germany; customer-environment deployment on request. RBAC, SSO, audit logs.
- Open APIs and SDK; customers own generated code, workflows and reports. Write-back limited to tickets, reports, plans.
AI Fact
- Frontier LLM agent: Claude Opus 4.6 / Sonnet 4.6 visible in the UI; enterprise page says any LLM.
- Grounding: company model + knowledge base + memory; the agent writes and runs code against real data and shows its steps.
- Customer ML models and simulations plug in as tools.
- 2024 origin was different: custom reinforcement-learning agents recommending machine setpoints (TechCrunch). The 2026 product leads with LLM agents.
Approach Fact
- Juna engineers map systems and flows; customer effort quoted at about 4 hours.
- Entry is a focused Proof of Value promising impact in weeks; expand use case by use case.
- Human-in-the-loop by design: customer sets objectives, constraints, approvals.
- Outputs flow into Teams, Slack, email, tickets and dashboards, so users stay in tools they already open.
Moat View
- Real: per-customer company model and knowledge graph create switching cost; heavy-process templates (paper, carbon black, asphalt, NH3 cold chain) and a 250-workflow library; German hosting + ISO 27001 for the Mittelstand; Kleiner Perkins money and Gartner listings for credibility.
- Thin: the reasoning layer is a licensed model anyone can buy; every input is someone else's system, so gaps in manual records stay gaps; public proof is unquantified (no named result, no metric on site or launch post).
Pricing Fact + View
- No public price. Enterprise commercial terms; data not used for model training.
- Demos sell in € impact (lost tonnes per day, €/month, kWe saved), never in seats.
- View: expect a paid PoV then a platform fee that grows per use case or site. Treat any number as unknown until a buyer shares one.
Market Fact
- Germany first; 2024 customers were all German (handful). Logos: Froneri (ice cream), Deutsche Gasrußwerke (carbon black), Currenta (chemical parks), Eratex (textiles).
- Started in energy-intensive heavy industry; launch post widens the target to all manufacturers.
- Gartner: key vendor, Industrial AI vertical agents (Innovation Guide for AI Agents, 2026); sample vendor, Hype Cycle for Discrete Manufacturing 2025.
- Narrative: European competitiveness through technology, retiring operators' know-how.
Juna vs Minder
Where the two products start, and where they meet.
| Juna | Minder | |
|---|---|---|
| Starting point | Existing systems: historian, MES, ERP, documents | First-party event capture from workers and signals; legacy systems as adapters |
| Primary user | Process engineers, plant managers | Frontline workers and managers on one event log |
| Input mode | Chat, uploads, connectors | Voice and vision first, native language, on devices in hand |
| AI stack | Claude agent + code execution + tool apps + memory | Agent loop over operational graph via read-only Trino, cited responses |
| Autonomy | Recommend, forward, write tickets/reports; human approves | Read-only/advisory in production; humans approve consequential actions |
| Proof shown | Logos and one CEO testimonial; no published metrics | Kasa: 7 departments, 100% assigned-worker adoption (founder-supplied, not independently audited) |
| Security | ISO 27001 certified, GDPR, German hosting | Controls in source; ISO 27001 / GDPR operations are milestones |
| Geography | Germany / DACH | Vietnam proving ground, Europe next |
Sources
- Product screens: paused frames from the product videos on juna.ai/foundation, juna.ai/chat, juna.ai/apps (captured 23 Oct 2026)
- juna.ai homepage — logos, testimonial, integration logos
- juna.ai/enterprise — deployment, security, commercial terms
- Agentic Factory OS launch post (Mar 2026)
- Juna on Gartner recognition (self-reported)
- TechCrunch, 18 Nov 2024 — founders, seed round, RL origin
- PitchBook profile — total raised, team size
- Silicon Canals — same seed reported as €7.1M
Conflicts left visible: founding year (2023 vs 2024); seed size in USD vs EUR. Customer logos are Juna's own claims and were not independently verified.