Competitor teardown, Celesnity product, 13 Jul 2026

Ethon: find the parameter that drives the loss, then steer it

Ethon builds a causal model of a production process. It shows which parameters drive a loss, recommends operating ranges with the expected gain, and keeps the line inside them. The product is narrow, consistent and easy to read on first use.

HQ / founded
Zurich
2021, spun out of ETH Zurich
Funding
$16.5M Series A
2024, led by Index Ventures; ~$23.5M total
Customers named
3+
Siemens, Roche, Lindt & Sprüngli (press)
Data needed
Variation
30 to 40 batches, about 10 parameters (Ethon FAQ)
Pricing
Not public
Enterprise sale
What matters for Minder
  • Ethon sells one loop: deviation, causal drivers, recommended range with a % gain, live control, rule monitoring. Every module shares one shell and one vocabulary.
  • The causal layer needs process data that varies: PLC, historian, quality results. Our target floors often have a cycle counter at best, so the detection layer does not transfer yet.
  • The interaction patterns do transfer: AI chat beside the causal graph, a projected gain on every recommendation, a current → recommended table with a live range bar, and a check that the line is stable afterwards.

Product shape

A process model in the middle, six workflows around it, one left nav for all of them.

SourcesProcess parameters, machine signals, quality results, unit and batch events, product images
→
Process modelProcess knowledge graph and causal model, picked from a Process Model selector
→
WorkflowsRoot cause, optimization, process control, anomaly detection, visual inspection, process mining; AI chat throughout
→
Human decidesRecommended ranges and setpoints, Adjust setpoint, rule monitoring to hold the gain

Nav seen on every screen: Dashboard, Unit Tracking, Insights, Data Visualization, Rule Monitoring, Anomaly Detection, Analysis, Optimization, Process Control, Process Mining, Value Stream Map, Cockpit.

ethon.com/ethonai-for-root-cause-analysis

Diagnose: root cause with the evidence in view

The strongest screen in the product. Diagnosis and explanation share one view.

Root cause analysis results

Root cause analysis

  • Process step importance by machine and an overall Pareto of parameters, both as estimated contribution to the target.
  • A causal graph from Parameter A through B and C to the target.
  • Chat on the right: "Parameter A usually moves ±2 units, now ±10", then cites internal documentation to suggest mechanical wear on Machine A.
Minder: answers in chat should sit beside the graph or chart they explain and cite the document behind the advice.
ethon.com/ethonai-for-process-optimization

Optimize and control

Recommendations carry their expected gain, then the control screen holds the line to them.

Loss optimization recommendations

Loss optimization

  • One card per parameter and machine: recommended range on a min–max bar, and the estimated change of the target, for example +17.97%.
  • Below, the response curve with the recommended band shaded, and outcomes inside vs outside the range.
Process control centerlining

Process control

  • Centerlining chart of the target with a simulation zone at the right edge.
  • Table per parameter: Out of Range or In Range, current → recommended value, and a live bar showing where the value sits in its band.
Minder: a clean template for any setpoint advice. It needs approve and verify states before it can touch a real line.
ethon.com/ethonai-for-early-fault-detection

Detect and inspect

Anomaly scores against a learned healthy state, and vision checks that feed back into the process.

Anomaly monitor

Early fault detection

Live anomaly score against a dashed threshold, the anomaly window shaded from 13 Jan, and the parameters with the largest influence listed below.

Visual inspection job

Visual inspection

Inspected image of a chocolate tray: 34 OK, 1 NOK, a heatmap on the faulty piece. Behind it, job yield (64.44%) and ROI scores against thresholds.

ethon.com/ethonai-for-production-traceability

Traceability: process mining on unit events

Unit and batch events rebuild the real path through the line.

Process mining graph

Process mining

  • Machines as nodes with average time; edges with transfer times, red where flow is slow.
  • Highlight loops, backward flow, skipped steps and unit assembly. Filter by variant, order or unit ID.
  • Feeds an automated value stream map.
Minder: closest Ethon feature to our event log. Our first-party events could draw this graph without a MES.
ethon.com/ethonai-for-plant-management

Role views

Seven role pages tell the same loop in each role's words. Two of them:

Plant management yield drift card

Plant management

OEE 88.1% split into availability, performance and quality. "Yield drift detected on Line 3", cause upstream temperature, action "Adjust setpoint T-14 to 162°C".

Frontline anomaly alert

Frontline

Issue detected, cause explained simply, action taken, stability restored. The operator sees one alert and one next step.

Marketing illustrations from the role pages, not product screens.

Analysis

Fact = from Ethon's site or press. View = our assessment.

Platform Fact

  • Six workflows: root cause, process optimization, process control, early fault detection, quality inspection, production traceability.
  • One connection to plant systems, data structured automatically, the same deployment pattern across sites.
  • Data scientists reach factory data and agent insights from Python and can add custom models.

AI Fact

  • Causal models of the process; Ethon argues publicly that random forests are wrong for root cause.
  • Chat grounded in the causal graph and internal documents.
  • Zero-shot product understanding for visual inspection.
  • Lineage: the older Analyst product ranked drivers by Pareto and relationship graphs.

Moat View

  • Real: causal modeling from ETH research; a narrow product with one clear loop; strong investors and blue-chip European customers.
  • Thin: depends on clean, varying process data; setpoint advice needs trust and often PLC access; little public outcome data beyond "up to 50% less waste" in press.

Approach Fact

  • Separate pages for management, plant management, manufacturing technology, process experts, frontline, data science and IT.
  • Rollout story: one standard layer across four factories.
  • Sells in loss and yield terms: OEE, scrap, waste.

Market Fact

  • Large discrete and process manufacturers in Europe and the US: Siemens, Roche, Lindt & Sprüngli; Bosch and IWC in one profile.
  • Demo data is generic (Machine A, Parameter K), so domain depth is hard to judge from outside.

For Minder View

  • Take: cause list beside the signal, quantified and dismissible recommendation cards, current → recommended table, verify after action.
  • Wait: anomaly scoring and causal discovery until a customer has varying process data.

Ethon vs Minder

Where the two products start, and where they meet.

EthonMinder
Starting pointProcess and quality data with enough variationFirst-party event capture from workers and signals; legacy systems as adapters
Primary userProcess experts, quality and plant managersFrontline workers and managers on one event log
Core objectDeviation, causal driver, setpointTask, decision and outcome on the plant model
AI stackCausal models, anomaly models, vision, chat over the graphAgent loop over the operational graph via read-only Trino, cited answers; world model in training
AutonomyRecommends setpoints; "Adjust setpoint" in marketing viewsRead-only and advisory in production; humans approve consequential actions
Proof shownNamed customers, press claim of up to 50% less wasteKasa: 7 departments, 100% assigned-worker adoption (founder-supplied)
GeographySwitzerland, Europe, USVietnam proving ground, Europe next

Sources

Conflicts left visible: Series A reported as CHF 15M, €15M, €15.3M and $16.5M, dated May 2024 by most sources; total raised $23.5M or $24.5M. Customer names come from press, not from Ethon case studies. Full research: SymphonyAI × Ethon screen atlas.