When inspector judgment becomes the variable nobody can spec away
Three shifts, three inspectors, three different calls on the same borderline defect – and a customer audit catches the inconsistency before you do. Fatigue and lighting don’t show up on a spec sheet, but they show up in your rejection rate.
Dumarey builds computer vision systems for defect, anomaly, and non-conformity detection directly in your production environment – on images and video, at the edge, on-premise, or in the cloud – replacing subjective borderline calls with a consistent, auditable threshold. This has been applied in textile and leather defect detection on premium manufacturing lines and in automotive component inspection with great results, such as false positive rate reduced by up to 98% (from 40 to 2%) for visual inspection for automotive injector production.

When the defect doesn’t show up until the product already left the plant
Some failure modes are invisible under normal light – a void inside a molded part, a thermal inconsistency in a casting, a material property that only deviates in a way a camera can’t see. By the time it surfaces, it’s a field failure or a warranty claim, not a line rejection.
For this, we go beyond standard imaging: hyperspectral and thermal imaging for material and process inspection, 3D reconstruction and monocular depth estimation for geometric non-conformities. This has been used for thermal monitoring of injection-molding processes and hyperspectral image analysis for product sorting – catching what visible-light inspection structurally cannot.

When the warning signs were in the data for weeks
A line goes down mid-shift. The post-mortem finds the signal was there in the sensor data days or weeks earlier – nobody was looking at the right variable, or the threshold that would have caught it wasn’t set.
Dumarey develops predictive maintenance and health-scoring models that detect anomalies and assign risk scores to components or systems before failures occur, enabling maintenance activities to be prioritised proactively rather than performed reactively. This approach has been successfully applied across several domains, including thermal management, measurement systems, and automotive after-treatment systems. By identifying recurring failure modes that conventional threshold-based alarms often fail to detect, these solutions have helped reduce vehicle downtime caused by after-treatment system issues by approximately 25%.

When root cause takes weeks because nobody can see the whole picture
An intermittent issue shows up in warranty claims months after the fact. Engineers manually cross-reference service logs, trying to reconstruct a pattern that’s scattered across systems that don’t talk to each other.
We use Bayesian networks, causal graphs, and probabilistic models to accelerate diagnosis when a failure mode is unclear – turning an investigation that used to take weeks of manual correlation into a structured decision process measured in days.

When fleet range planning is still built on averages
A charging schedule built around an “average” driver doesn’t hold up against real driver behavior, real degradation curves, and real route variance – and the gap shows up as missed SLAs or underused charging infrastructure.
For electrified fleets, we build predictive models for driver and charging behavior, battery thermal intelligence, and route-energy optimization, replacing averages with models grounded in how your specific fleet actually operates.

When the person who knows where everything is decides to retire
The senior engineer who’s been there twenty years is the actual search index for your organization’s knowledge. FMEAs, lessons learned, simulation results – technically documented somewhere, practically retrievable only by asking them. When they leave, that retrieval capability leaves with them.
Dumarey builds knowledge management systems and retrieval-augmented generation (RAG) over internal documentation, plus agentic workflow automation for repetitive tasks, turning tribal knowledge into something the next engineer can actually query.

When RFQ season turns your best engineers into document processors
A stack of RFQs (Request For Quote) lands with a two-week turnaround, requirements buried across scattered PDFs and spec sheets, and the proposal team spends the first three days just extracting what’s actually being asked for before anyone can start drafting a response.
We build document intelligence pipelines for RFQ scanning, proposal drafting, and structured data extraction – compressing the part of the cycle that’s pure document processing so your engineers spend their time on the part that actually requires judgment.

When legacy code and content work both need the same underlying capability
The same generative AI foundation that handles document intelligence also extends to engineering productivity – including embedded-software reverse engineering, where GenAI reconstructs documentation from legacy code – and to lighter-touch applications like marketing content generation, sales lead qualification and pricing intelligence, and technology-trend mapping for consortium building and funding applications.