Real AI and computer-vision systems we have designed and shipped for businesses across the Benelux — from precast concrete lines and low-emission zones to cold-chain fleets and clinical workflows. No jargon, no vapourware: just systems built to run in production.
Where you see a result tagged Industry benchmark, that figure comes from peer-reviewed studies or sector reporting for the underlying technology — not from our own client deployments. We show it to set realistic expectations for what the approach can achieve. The figures we report for your project always come from a field trial on your own data and operations. Cases tagged Client result or Live deployment are real Crux Digits engagements, with the client name withheld until we have written permission to publish it.
Each card jumps to the full write-up: the problem, how the system works, the honest results framing, and a short FAQ.
Catch cracks, spalling, chipping and voids on precast segments automatically, right on the production line.
In short →Read plates and check zone eligibility in real time — rain, glare or motion blur — with an audit trail behind every decision.
In short →Turn ordinary road footage into a live, prioritised map of potholes, cracks and obstacles — so teams fix the right defects first.
In short →Tell a real temperature excursion apart from normal fluctuation — and alert the team early enough to save the load.
In short →Read ECGs and scanned reports, extract the clinically relevant features, show the reasoning — and keep the clinician in control.
In short →Adaptive questionnaire screening that helps clinicians focus the right questions and surfaces a structured likelihood summary.
In short →A live occupational-health platform that speeds up and standardises medical-fitness evaluations — already processing assessments at national scale.
In short →Turn demand signals into a production plan, smooth mould-change losses and defend market position — for a €235M footwear manufacturer.
In short →Find at-risk customers and the levers to keep them — so marketing targets the right cohorts instead of spraying spend.
In short →Spot the prized two-leaves-and-a-bud growth, catch pests and disease early, and put crews where they matter most — across 3,200 ha.
In short →Learn the signatures of failure and flag them early — so maintenance happens on schedule, not in a crisis.
In short →Draft and clarify discharge summaries fast — cutting turnaround and documentation-error risk while the clinician keeps sign-off.
In short →Watch ad positions and adjust bids in real time to hit target placements within a set budget — no manual bid-watching.
In short →Precast is only as reliable as its inspection. A defective segment that reaches a tunnel is expensive — or dangerous — to replace. We built a vision system that catches cracks, spalling, chipping and voids the same way, every time.
We build AI vision systems that detect and classify damage on precast concrete — cracks, spalling, chipping, voids — directly on the production line or during tunnel inspection. Detection and segmentation models (YOLO, U-Net, Mask R-CNN, Vision Transformers) are adapted for construction materials and run at the edge, so the line never waits and your inspectors review only the uncertain cases.
Manual inspection is slow, subjective and hard to scale. Inspectors judge severity differently, casting-yard light is uneven, and dust, form-release agent and water staining trigger constant false alarms. The client needed inspection that was objective, traceable and fast enough to keep pace with production — and could later feed a robotic cell.
Real-time intake from line cameras and 3D/LiDAR, aligned by a synchronisation framework.
De-noising, illumination and geometry correction, with optional visual + LiDAR fusion to kill false positives.
Cracks segmented at pixel level; spalling, chipping and voids categorised by likely severity.
Results pushed via API to robotic manipulators; low-confidence cases flagged for human review.
Benchmark basis: peer-reviewed studies, 2023–2025 (e.g. YOLOv8 concrete-crack: 91.8% precision / 92.5% recall; SDNET2018 ensembles up to ~98%). These are sector benchmarks for the technology — not Crux Digits’ own results. For your project we report the verified precision and recall from a field trial on your own segments.
With a controlled line camera and a model retrained on your own segments, detection of fine-width cracks is reliable. We report precision and recall from a field-trial baseline on your material — not generic benchmarks.
Yes. Inference runs at the edge, so detection keeps pace with line speed and can trigger robotic handling decisions immediately.
No. It standardises detection and flags low-confidence cases for review, so inspectors spend their time on judgement, not repetitive scanning.
A low-emission zone only works if enforcement is reliable and fair. Manual checks do not scale; brittle cameras fail in rain and glare. We built a system that reads plates and verifies eligibility in real time — with an audit trail behind every decision.
Computer vision (YOLOv8) detects vehicles and plates in rain, glare and motion blur, OCR reads the plate, and a rule engine cross-references registration data to confirm emission category or eligibility — flagging violations and logging an auditable record, with role-based dashboards for operators.
Manual enforcement cannot cover a zone around the clock, and off-the-shelf cameras drop accuracy the moment conditions get hard: night, rain, low sun, partial occlusion, motion blur. The city needed enforcement that was consistent, fair and fully auditable — every decision backed by evidence.
A fine-tuned YOLOv8 finds vehicles and plates through occlusion, motion blur and variable light.
Contrast, binarisation and skew correction, then OCR with plate-format and regional rules.
Plates validated against registration APIs for make, year, emission category and ownership.
A rule engine evaluates eligibility, flags violations and logs everything with role-based access.
Benchmark basis: ANPR industry reporting and studies — real-world accuracy typically 90–98%, capture/read rates ≈98%/95%. These are sector benchmarks, not Crux Digits’ own deployment figures. We report verified read-accuracy from your own roadside footage. Built as a proof of concept for the Municipality of Antwerpen (District Antwerpen).
A YOLOv8 detector plus OCR preprocessing handles rain, glare, motion blur and partial occlusion; we report read-accuracy from your own roadside footage, not lab conditions.
Plate data is personal data. We design with data minimisation, retention limits and access control; the legal basis and retention policy are set by the operating authority.
Yes. Structured results and an auditable log are exposed via API, so violations and evidence drop straight into the systems your operators already use.
Roads degrade continuously, but inspection budgets are fixed. We built a system that turns ordinary road footage into a live, prioritised map of potholes, cracks and obstacles — so teams fix the right defects first.
Semantic segmentation finds cracks and surface wear; object detection (YOLOv8) finds potholes and debris; temporal filtering across frames cuts false positives; and every defect is mapped to GPS coordinates and ranked by severity in dashboards built for road authorities.
Periodic manual surveys are expensive and always slightly out of date, while single-frame detectors mistake shadows and wet patches for damage. The authority needed continuous, prioritised visibility of the network — without commissioning a fleet of specialist survey vehicles.
Multi-view video with GPS sync, then stabilisation, deblurring and colour normalisation.
Segmentation for cracks and wear; YOLOv8 for potholes and debris, trained on dust and glare.
Consecutive-frame comparison and optical flow suppress shadows and wet patches.
Each defect is pinned to GPS; dashboards show heatmaps and maintenance priorities.
Benchmark basis: peer-reviewed road-defect studies, 2022–2025 (YOLOv8 pothole mAP ≈81–86%; tuned models to ~95%; ~20 FPS real-time). These are sector benchmarks, not Crux Digits’ own route results. For your network we report mAP and false-positive rates from a route trial on your own footage.
Temporal filtering — consecutive-frame comparison and optical flow — plus augmentation on dust, glare and motion suppress the artefacts that single-frame detectors mistake for defects.
Vehicle-mounted or roadside cameras with GPS sync — no specialist survey vehicle required.
Each defect is graded by severity and pinned to GPS, so dashboards surface the highest-risk repairs first instead of a flat list.
One undetected temperature excursion can write off a whole pharma or food shipment. We built an AI-plus-IoT system that tells a real anomaly apart from normal fluctuation — and alerts the team early enough to act.
Real-time streams of temperature, humidity, vibration and GPS feed anomaly-detection models that separate true deviations from normal variation, link readings to route and environment, and trigger alerts before a breach spoils a load — with dashboards for live fleet status and historical analysis. Edge buffering keeps it working when connectivity drops.
Threshold alarms either miss slow drifts or cry wolf at every door opening, so teams learn to ignore them. The operator needed early, trustworthy warnings that distinguished a genuine excursion from harmless variation — and kept working through the connectivity dead zones on a route.
Stream temperature, humidity, vibration and GPS, with edge buffering for dead zones.
Anomaly models flag deviations; temporal smoothing separates noise from real drift.
Automated alerts for breaches and predicted deviations — before they happen.
Live fleet dashboards plus historical trends for audits and SLAs.
Benchmark basis: cold-chain and UN/FAO reporting, 2024–2026 — real-time IoT monitoring is associated with up to ~30% spoilage reduction, and roughly 30% of food waste is preventable through a better cold chain. These are sector benchmarks, not Crux Digits’ own fleet results. For your fleet we report verified spoilage and alert-accuracy figures from a live trial.
Temporal smoothing and contextual analysis distinguish normal variation — a door opening, a gradient — from true anomalies, so alerts fire on real risk instead of every fluctuation.
Yes. Edge buffering stores readings during outages and syncs when the connection returns.
Yes. The models project the temperature trajectory and warn when a load is heading for an excursion, giving the team time to intervene.
Cardiology generates more ECGs and scanned reports than clinicians can review at leisure. We built decision support that reads them, extracts the clinically relevant features and shows its reasoning — so clinicians spend time on judgement, not transcription.
Vision models segment multi-lead ECGs and extract waveform features (P-wave, QRS, T-wave), combine predictions with rule-based clinical heuristics for explainable, lead-wise reasoning, score scan quality, and flag low-confidence cases for clinician review. It supports clinicians — it does not replace them.
High volumes of ECGs and scanned reports arrive faster than they can be read carefully, and a black-box label is no use to a clinician who needs to justify a decision. The team needed support that was fast, explainable, and honest about its own uncertainty.
Scanned PDFs, ECG images and report photos; deskew, denoise, normalise; segment leads.
Vision models pull P/QRS/T morphology and flag possible arrhythmia or ischaemia.
AI predictions combined with clinical heuristics for explainable, lead-wise output.
Structured findings with confidence scores; low-confidence cases go to a clinician.
Benchmark basis: peer-reviewed cardiology studies, 2024–2025 (deep-learning ECG models reach cardiologist-level accuracy ~88–94%; AUC up to ~0.99 on large datasets). These are sector benchmarks for decision-support models — not Crux Digits’ own validation. Any clinical claim must come from your own validated study. Decision support only; not an autonomous diagnosis, and clinical use requires appropriate regulatory clearance (e.g. EU MDR) and clinician oversight.
No. It is decision support: it extracts features and suggests findings with confidence scores and reasoning, and routes uncertain cases to a clinician.
Yes. It combines model predictions with rule-based clinical heuristics to give lead-wise reasoning, not a black-box label.
Clinical deployment requires the appropriate clearance (for example EU MDR) and clinician oversight; the system is built to support that process, not to bypass it.
Early signs of Parkinson’s and Parkinsonian syndromes are easy to miss in a busy clinic. We built an adaptive, questionnaire-based screening tool that helps clinicians focus the right questions and surfaces a structured likelihood summary — support for the preliminary evaluation, not a diagnosis.
A dynamic, condition-driven questionnaire adapts to each response, collecting only what is relevant. AI classifies the answers into defined likelihood categories and produces a consolidated, clinician-friendly summary to support — not replace — clinical judgement. It is privacy-first by design: no personally identifiable information, with age the only demographic used for context.
Early Parkinsonian signs are subtle and easy to miss in a short consultation, while a rigid questionnaire wastes time on irrelevant items. The clinic needed a tool that adapted to each patient, respected privacy, and produced something a clinician could actually act on — without ever implying a diagnosis.
A condition-driven questionnaire adapts in real time, asking only the relevant follow-ups.
No personally identifiable information is collected; age is the only demographic used for context.
AI maps responses to predefined likelihood levels for the relevant conditions.
A consolidated summary supports clinical interpretation — never a definitive diagnosis.
Benchmark basis: peer-reviewed neurology studies, 2018–2026 — AI Parkinson’s detection spans ~78–96% accuracy across modalities, with questionnaire-based ML (MDS-UPDRS) models reaching >95% accuracy and AUC. These are sector benchmarks, not Neuro Path Finder’s own validated metrics. Developed with input from a board-certified neurologist specialising in movement disorders. Screening-support only — it produces an indicative likelihood summary, not a diagnosis, and clinical use requires clinician oversight and appropriate regulatory clearance.
No. It classifies questionnaire responses into likelihood levels and produces a summary to support a clinician’s preliminary evaluation — the diagnosis remains the clinician’s.
It is privacy-first by design: no personally identifiable information is collected, and age is the only demographic parameter used to add clinical context.
It is condition-driven: each answer shapes the next set of questions, so the tool only collects what is clinically relevant for that patient.
Large-scale occupational-health screening is high-volume, repetitive and easy to apply inconsistently. AMAN AI is a live platform that speeds up and standardises medical-fitness evaluations while keeping clinicians firmly in charge — already processing assessments at national scale.
AMAN AI is a customised suite of AI models and agents that runs inside existing clinical workflows to make medical-fitness evaluations faster, more consistent and guideline-aligned. It combines three decision-support models — a Classification model that consolidates multi-source medical data, a cardiovascular-risk model that flags latent risk from biomarker correlations, and a mental-fitness model — under a clinician-led "maker-checker" approach.
National-scale fitness screening is high-volume and repetitive, which makes consistency hard and latent risk easy to overlook. The provider needed to speed evaluations up and standardise them without taking the final decision out of clinicians’ hands — and without ripping out the workflows their teams already use.
Extracts and consolidates clinical parameters from unstructured data, correlates anomalies with guideline criteria, and proposes an indicative fitness classification the clinician can accept, modify or override.
Machine learning on vital signs and biomarkers surfaces latent cardiac risk — even in people who look normal on standard indicators like ECG — as an interpretable, explainable risk score.
Analyses physiological and biological markers to produce an indicative Mental Fitness Score that supplements established fitness guidelines.
Live deployment: delivered by Crux Digits and partners at a national occupational-health provider, in production since February 2025, supporting large-scale medical-fitness evaluations for contractor recruitment and onboarding in the energy sector. Client name withheld pending permission. A clinician-led maker-checker model keeps every fitness, cardiovascular and mental-health output indicative — supporting, never replacing, the clinician’s final decision; deployment requires appropriate regulatory clearance.
No. It proposes an indicative classification that the clinician can accept, modify or override — a controlled maker-checker approach that keeps decisions clinician-led while adding systematic AI validation.
It learns correlations among indirect biomarkers and longitudinal physiological patterns, identifying latent risk that may not be visible through standard indicators alone, and reports it as an explainable risk score.
Yes. It is a customised suite of models and agents designed to operate inside existing workflows, aligned with established occupational-health guidelines.
A leading footwear manufacturer (€235M turnover) was wrestling with complex day-to-day production planning and resource loss during mould changes. We built forecasting that turns demand signals into a production plan — and protects market position.
Time-series and gradient-boosting models (Prophet + XGBoost) forecast demand and feed day-to-day production planning, smoothing mould-change losses and aligning capacity to real demand.
Day-to-day planning across a wide SKU range was largely manual, and every mould change leaked capacity. Without a reliable demand signal, the manufacturer was over- and under-producing in turn — and losing ground to aggressive competitors.
Prophet models seasonality and trend in demand across SKUs.
XGBoost layers in drivers and corrects the forecast on fresh data.
Forecasts drive production scheduling and reduce mould-change waste.
Real result from a Crux Digits engagement; client name withheld pending permission.
Prophet captures seasonality and trend cleanly; XGBoost layers in the additional drivers and corrects the forecast as fresh data arrives — together they are more accurate than either alone.
A reliable demand signal lets the plant sequence production so mould changes line up with real need, instead of reacting to surprises.
A major spice exporter faced rising churn from shifting market preferences and campaigns that were not landing. We built churn prediction that finds at-risk customers and the levers to keep them.
Decision-tree and K-means models segment customers and score churn risk, so marketing can target the right cohorts with the right campaigns instead of spraying spend.
Retention spend was spread evenly across the whole base, so campaigns were expensive and underwhelming. The exporter needed to know which customers were actually about to leave — and what would make them stay — before the budget was gone.
K-means groups customers by behaviour and value.
Decision trees score churn risk and surface the drivers.
Tailored campaigns aim at the cohorts most worth keeping.
Real result from a Crux Digits engagement; client name withheld pending permission.
Decision-tree models score each customer on churn risk and surface the behavioural drivers behind that score, so the signal is actionable, not just a number.
K-means groups customers by behaviour and value so campaigns can be tailored to each cohort, concentrating retention spend where it returns the most.
A large tea-estate group (3,200 ha across 7 estates, 10,000+ field workers) needed to spot the prized two-leaves-and-a-bud growth, catch pests and disease early, and put crews where they mattered. We built vision that does exactly that.
A convolutional vision model (VGG-19) identifies the high-value two-leaves-and-a-bud composition and flags potential pest and disease issues early, so workforce and plant-protection spend go where they earn their keep.
Across thousands of hectares and a 10,000-strong workforce, quality and plant-protection decisions relied on manual spot-checks and experience. Effort and chemicals were spread too thinly to be efficient, and early pest or disease signs were easy to miss until they spread.
VGG-19 grades leaf composition and spots early pest / disease signs.
High-quality and at-risk areas are mapped across the estates.
Crews and plant-protection effort follow the map, not guesswork.
Real result from a Crux Digits engagement; client name withheld pending permission.
The model identifies the high-value two-leaves-and-a-bud composition and maps where it is concentrated, so harvesting effort focuses on the most valuable growth.
By mapping at-risk areas precisely, crews and chemicals are applied where they are actually needed instead of blanket-treating every estate.
A sugar producer was losing efficiency to recurring, unplanned downtime. We built models that learn the signatures of failure and flag them early — so maintenance happens on schedule, not in a crisis.
Convolutional models (CNNs) learn the patterns that precede equipment failure from sensor and inspection data, flagging issues early so teams intervene before a breakdown — turning unplanned downtime into planned maintenance.
Run-to-failure maintenance meant recurring, unplanned stoppages that cost output and forced expensive emergency repairs. The producer needed early warning of failures so work could be scheduled into planned windows instead of triggered by a breakdown.
CNNs learn normal vs. pre-failure signatures from plant data.
Early anomalies trigger an alert before failure occurs.
Maintenance is planned into the window, not forced by a breakdown.
Real result from a Crux Digits engagement; client name withheld pending permission.
Sensor and inspection data from the equipment; the CNN learns the difference between normal operation and the patterns that precede a failure.
Early, specific warnings let teams plan the right intervention in advance, so work is targeted and scheduled rather than diagnosed under pressure during a breakdown.
A hospital group was burning clinician time reading and interpreting discharge summaries, with delays and a real risk of documentation errors. We built NLP that drafts and clarifies them, fast.
NLP models read clinical inputs, draft a clear, structured discharge summary and cut the manual reading and interpretation that slowed clinicians down — reducing both turnaround and documentation-error risk.
Producing and interpreting discharge summaries by hand was slow and error-prone, holding up the discharge process and pulling clinicians away from care. The group needed faster, clearer documentation without compromising the clinician’s final review.
NLP parses the clinical record and key parameters.
A clear, structured discharge summary is generated for review.
The clinician checks and signs off — faster, with fewer errors.
Real result from a Crux Digits engagement; client name withheld pending permission. Decision support only — the clinician retains sign-off on every summary.
It drafts a clear, structured summary, but the clinician always reviews and signs off — the system speeds the work up without removing clinical control.
By extracting and structuring the record consistently, it removes much of the manual transcription where errors and omissions creep in.
A Benelux seller on Bol.com wanted better product visibility without overpaying for ads. We built a system that watches ad positions and adjusts bids in real time to hit target placements within a set budget.
The system reads current ad positions via the Bol.com Advertising API and adjusts bids dynamically against the seller’s max-bid and target pages — automatically defending the placements that drive sales, without manual bid-watching.
Marketplace ad positions shift constantly as competitors change their bids, so manual bidding either overpays to stay visible or quietly loses placement. The seller wanted to hold target positions within a fixed budget — without watching bids all day.
Set a maximum bid and the target product pages to optimise.
The system retrieves live ad positions and the competition.
Bids update automatically via the Bol.com Advertising API.
Real Crux Digits engagement; client name withheld pending permission. Capability described as delivered — no published outcome metric is claimed.
It reads live ad positions via the Bol.com Advertising API and adjusts bids toward your target placement, always staying inside the maximum bid you set.
Yes. You configure the maximum bid and the target product pages; the system optimises within those limits and never bids past them.
Tell us what you are trying to fix. We will scope a focused pilot on your own data and report real numbers — not promises.
Book a free consultation →