FAQ
Frequently asked questions
Do you need images of defective parts to train the AI ?
No. DeepHawk learns what a good part looks like and flags anything that deviates from it. A model is trained on roughly 30 to 50 images of conforming parts. There is no defect library to build, no labelling campaign, no waiting months to collect enough failures.
This is the main difference with traditional deep-learning approaches, which typically need thousands of defective samples before they produce anything usable.
Can it catch defects nobody defined in advance?
We offer a [brief explanation of how your product or service helps your customers]. Our process is [simple/efficient/convenient] and [positive adjective] for you.Yes, and this is one of the real advantages of an anomaly-detection approach.
Each model learns what normal looks like in a given zone and flags anything that deviates, including issues no one specified upfront. In practice the system regularly surfaces problems that were never on the inspection checklist. The one condition is that a zone must be covered by a model: DeepHawk inspects the areas you configure.
Do we have to replace our cameras, or buy smart cameras?
No. DeepHawk works with standard industrial cameras (USB, GigE, MIPI), as well as X-ray, microscopy, thermal and hyperspectral imaging. You do not need smart cameras, because the intelligence sits in the software rather than in the camera, which also makes the hardware cheaper. If no camera is in place yet, we write the vision-system specification so your integrator can source the right one.
Resolution is driven by the smallest defect you need to catch, and the algorithm adapts to whatever resolution you provide.
Does this only work on surface defects on simple parts, or on assembled equipment too?
Assembled equipment as well. Typical deployments include solder joints on electronic boards and engine wiring harnesses, not just surface finish on a single machined part.
Complex assemblies are also where positional errors matter, and the system detects them: a connector fitted at 90 degrees in the wrong orientation, for example, is picked up as an anomaly like any other deviation from the good part. The fastest way to settle the question for your own products is a test on your images.
Can our own team retrain the models when quality criteria change or a new product arrives?
Yes, and that is the point. Your quality engineers and operators add a new reference or adjust criteria themselves by uploading new conforming images, adding or removing inspection zones and tuning sensitivity per zone.
During deployment we train your operators and quality engineers to full autonomy on model creation and configuration. You do not call us back every time your quality plan evolves.
How do you keep false alarms under control? Does the system give a confidence level?
The model produces a continuous anomaly score, not a simple pass or fail. The stronger the anomaly, the higher the value. You can define several decision bands, for example a “check” range that triggers human review and a “not OK” range beyond an upper threshold, and every threshold is configurable.
We start deployments with a conservative setting that generates few false alarms, then tighten it as operator feedback comes in. One client moved to a 1% false-alarm rate within three days.
What does DeepHawk need from our IT, and where does our data go ?
DeepHawk runs on a standard industrial PC under Windows. It is a frugal AI, built to run on the standard industrial hardware you already have, so it drops into your existing environment without a dedicated compute server.
Everything runs on premises: images, models and results stay inside your plant, nothing is sent to the cloud and nothing is sent to us. That is what makes the solution deployable in aerospace, defence and any environment where part geometry is sensitive intellectual property.
What is the smallest defect you can detect ?
Down to pixel level. Because the model learns normality instead of a catalogue of defect shapes, there is no minimum defect size built into the training set. The practical limit is your camera resolution: we convert a target defect size into pixels based on camera and distance parameters.
This is anomaly detection rather than dimensional metrology. You then decide what counts as a defect, since sensitivity, maximum defect size and maximum total defect surface are set region by region according to your own quality plan.
DeepHawk is reinventing AI and uplifting machine learning capabilities, operational performance and frictionless deployment to a brand new level.
location
© 2026. All rights reserved.
Headquarters
2B rue de la Châtaigneraie
35510 Cesson-Sévigné
France
DeepHawk Inc.
17940 Farmington Rd
Suite 216B
1003 Livonia, MI 48152
USA