Custom computer vision that really works in your production environment, warehouse, shop or field. From object detection and image recognition to OCR, segmentation and anomaly detection, built as an integration into your existing software, not as a standalone proof of concept.
Cameras can now see more than people can keep track of. On an assembly line, more product passes by in an hour than a QA inspector can visually check. In a warehouse, thousands of labels are scanned. On a construction site, someone always has to check whether helmets are being worn. Computer vision (CV) automates these kinds of tasks — not as magic, but as a reliable part of your operational software.
We build computer vision applications that slot straight into your existing stack: a YOLO model running on an edge device next to the production line, an OCR pipeline that reads invoices and posts the lines into your accounts, or a classification model that pre-filters insurance claims for your back office. Not a standalone "AI pilot" that never goes live, but a working integration with your ERP, WMS or CRM. For strategic planning, we also help with a broader AI strategy in which computer vision is one of the applications.
In recent years the tooling has matured quickly. Pre-trained models such as YOLO, Detectron2, Segment Anything and Vision Transformers have taken away much of the heavy lifting: you no longer need to label millions of images, but a few hundred to a few thousand per class. At the same time, edge devices have become powerful enough to run those same models locally — next to the production line, in a drone or in a barrier installation. And multimodal LLMs such as GPT-4o-vision, Claude Vision and Gemini Vision give you a new, shorter route for zero-shot and few-shot tasks for which you would previously have trained your own model. Which route fits depends on your data, your latency requirements, your privacy context and the business case around it — and that is exactly what the first conversation is about.
Depending on what you need to recognise, how critical it is, and whether existing models are good enough. In the first conversation we advise which type fits.
For standard tasks — general object detection, OCR on regular documents, face blurring for GDPR compliance — we use Google Vision, Azure Computer Vision, AWS Rekognition or an open-source pre-trained model. Fast implementation, no proprietary training set required, pay per call.
When you need to recognise something specific — your own products, defects nobody else spots, industry-specific documents — we fine-tune a pre-trained model (YOLOv8, Detectron2, Vision Transformer) on your imagery. Includes the labelling process, training pipeline and inference service.
For applications with low-latency requirements — live production inspection, perimeter surveillance, vehicle tracking — the model runs on an edge device (NVIDIA Jetson, Google Coral, Hailo) or within a real-time video pipeline. We build the entire stack: capture, pre-processing, inference, post-processing and dashboard.
A working CV system you can maintain yourself, plus everything around it to build further.
Computer vision touches many sectors. Below are the scenarios where we support clients. If you recognise one of them, we would be happy to talk further.
Cameras above the assembly line detect scratches, dents, missing components or incorrect colours. A YOLO or anomaly detection model (PatchCore, PaDiM) flags defective units so they are sorted out automatically before your QA team even sees them.
Parcels and pallets moving through your warehouse, labels and barcodes on damaged boxes, and automatic counting of stock on shelving. We combine OCR (PaddleOCR, Google Vision) with object detection and integrate it with your WMS.
A claimant uploads a photo, a classification model assesses the type and severity of the damage, and the claim is pre-sorted. The same pattern works for property inspections, drone photos and pre-acceptance. It fits well with custom LLM integrations, where text and images are analysed together.
Shelf cameras recognise out-of-stock situations and planogram deviations. In agriculture we do similar work with drone imagery: crop stress, fruit ripeness and livestock counts in a field, using segmentation on satellite or drone photos.
Detection of helmets, high-visibility vests and safety goggles on construction sites, combined with perimeter monitoring. We only work on applications that are ethically and legally sound: no biometric mass surveillance, no emotion recognition, but safety within defined contexts.
ANPR for car parks, charging points or barriers. Vehicle tracking for traffic analysis and occupancy. Licence plate data is personal data under the GDPR, so we build in appropriate retention and storage periods.
For applications involving medical imaging, we work only via an MDR-certified route, often together with a medical device partner. For non-diagnostic tasks (workflow support, anonymisation, image routing) we build in-house to NEN 7510 requirements.
Extracting text and structured data from photos or scans, with validation against your master data. We combine PaddleOCR or Tesseract with multimodal LLMs for the difficult edge cases where rule-based OCR falls short.
With OneDayBuild, we turn your idea into something tangible in one day for €1,150, so you can see whether further development is worth the investment. Decide to go ahead with the full build? Then we credit the full cost.
Explore OneDayBuild →A brief tour of the models, frameworks and deployment targets we use in practice. We choose per project; there is no one-size-fits-all stack.
YOLOv8, YOLOv9 and YOLOv10 for real-time detection. Detectron2 and MMDetection for heavier tasks. Segment Anything (SAM) for pixel-precise boundaries. Vision Transformers (ViT, DINOv2) where the transformer architecture offers an advantage.
For zero-shot tasks, complex reasoning on images, or where text and images need to be analysed together. Often a quick way to validate whether CV adds any value at all before we train a custom model.
PyTorch as the training framework, ONNX for cross-platform export, OpenCV for pre- and post-processing. Tracking with SORT, ByteTrack or DeepSORT. Anomaly detection with PatchCore or PaDiM for industrial inspection.
Cloud GPUs for batch processing and training, NVIDIA Jetson and Google Coral for edge, CoreML and TFLite for mobile and on-device. Inference is packaged as a REST, gRPC or WebSocket service, suitable for both real-time and batch use.
For labelling we work with Label Studio, CVAT or V7. We track experiments and model versions in Weights & Biases or MLflow. CI/CD pipelines for training and deployment, plus monitoring of precision, recall and drift.
When you have too few production images, we generate synthetic data: Stable Diffusion for variation, or a 3D renderer for pixel-precise labels. This is often the key difference between "it works in the demo" and "it works in production".
A conversation in which we sharpen the use case: what needs to be recognised, how often, how quickly, and with what accuracy it becomes valuable. We often carry out a short feasibility scan to determine whether an existing cloud model is sufficient or whether training is needed.
Almost every CV project stands or falls with the data. We help collect representative images, choose the right labelling platform (Label Studio, CVAT, V7) and agree an acceptance criterion for dataset quality.
We fine-tune a pre-trained backbone on your data, track experiments in Weights & Biases or MLflow and validate against a hold-out set. You receive confusion matrices, precision/recall charts and concrete examples of what the model does and does not see, so you know what you are buying.
The model is packaged as an inference service (REST, gRPC or a converted CoreML/TFLite file for edge) and integrated with your existing software. We work in sprints and deliver something you can test every two weeks, with monitoring built in from day one.
After go-live, we monitor for model drift, retrain periodically on new data and, if you wish, offer a maintenance contract under which we evolve the model alongside your operations.
What clients usually want to know before we start.
A half-hour introductory call, no obligation. We listen to the use case, ask about your data and give direction on feasibility, approach and which models fit. If it sits within a broader AI landscape, we'll extend the conversation to generative AI solutions or a wider strategy.
Appfront uses cookies and similar technologies to keep the website working properly, for analytics and for marketing. You choose what you allow. Read more in our privacy policy.
