Crackz
Crackz is AI defect-detection software that runs pixel-level semantic segmentation to find cracks, corrosion, spalling and weld flaws across harbour piers, bridges, factory lines and energy assets. It ships as a desktop app, a web interface and a command-line tool.
What is Crackz?
Crackz is defect-detection software built around deep-learning semantic segmentation. Rather than labelling a photograph as cracked or not, it classifies every pixel, which lets it outline the shape and extent of cracks, corrosion, spalling, weld defects, casting porosity and paint failures. It was originally built for concrete deterioration on harbour piers, docks, bridges and seawalls, and has since been applied to manufacturing quality assurance on production lines and to industrial and energy asset integrity in chemical plants, refineries, wind farms and rail.
The software ships in three forms from a single commercial source-available codebase: a desktop application, a web interface and a command-line tool, with Docker distribution. The desktop application is organised around four stages — project configuration, mask editing, training and detection. Everything is reproducible from one YAML file covering model configuration, training schedule, detection parameters and data paths. The published stack is PySide6 for the interface, PyTorch Lightning for training, Streamlit for the web view, and SegFormer encoders B0 to B5, with SegFormer-B2 as the default. Input tiles are 512 by 512 pixels, the detection threshold is 0.50, and the loss combines Dice and binary cross-entropy. It runs on CUDA, ROCm or CPU.
Two adoption paths are offered. Pre-trained checkpoints cover standard crack detection on concrete and steel and require no training data to start. Custom training targets a specific domain and unusual defect types, needing roughly 100 images for a pilot and 500 or more for production use, with an active-learning loop to reduce annotation cost. Checkpoints record their own annotation schema, so incompatible models are flagged automatically.
Output is designed for downstream use: annotated PNGs, GeoJSON ready for QGIS or other GIS pipelines, severity buckets, per-pixel and per-image confidence, and audit-grade Word or PDF reports. The publisher is a very small Stockholm engineering studio that also sells senior consulting, and presents Crackz as its flagship product at version 0.81. Two neighbouring efforts exist: Survey Ops, a marine survey data catalogue, and AI Slam, described explicitly as an architecture still being built rather than a shipping platform.
What it does
- Detect cracks, corrosion, spalling and weld defects at pixel level using semantic segmentation
- Process thousands of inspection images in a single run instead of reviewing them by hand
- Rank findings into High, Medium, Low and Clean severity buckets with per-pixel and per-image confidence
- Train a custom model on your own multi-class defect taxonomy, or deploy a pre-trained checkpoint
- Annotate imagery in a brush-based mask editor with propagation across image series
- Export annotated PNGs and GeoJSON straight into QGIS or another GIS pipeline
- Produce audit-grade Word and PDF reports with GPS and EXIF traceability
When to use Crackz / When not to
A quick filter to help you decide if Crackz is the right fit.
When to use Crackz
- Harbour, bridge and seawall inspection programmes dealing with concrete and steel deterioration
- Manufacturing QA teams looking for surface defects, weld flaws, casting porosity or paint failures on production lines
- Asset-integrity engineers in chemical, refinery, wind and rail operations who need condition-based maintenance
- Organisations that must keep imagery on their own hardware, since the software can be deployed on-premise
- Technical teams comfortable with a command line and a GPU, who want to train a model on their own defect taxonomy
When not to use Crackz
- Buyers who need a published price, because every offer on the site says only "Contact us for pricing"
- Teams that require contractual guarantees up front: the publisher puts no terms of service, no privacy policy and no DPA online
- Developers looking to call a hosted API, as the product exposes neither an API nor any API documentation
- Anyone wanting to sign up and try the tool alone, since access runs entirely through a sales conversation
- Mobile-first users, as no iOS or Android application exists
How to use Crackz
A typical end-to-end flow, from setup to results.
- Contact the publisher, since there is no self-service sign-up: every button leads to a demo request or a sales conversation
- Install the software on your own hardware as a desktop application, a command-line tool or a Docker container
- Decide between a pre-trained checkpoint for standard crack detection and a custom model for your own defect taxonomy
- Import your inspection imagery, for example with "crackz import-images --src ./photos"; raw, train, validation and test splits are created automatically and EXIF data is preserved
- If training a custom model, annotate images in the mask editor via Tools then Mask Editor, using the brush and propagating masks across image series
- Set the project up in a single YAML file covering model configuration, training schedule, detection parameters and data paths
- Launch training, for example "crackz train --epochs 20", and follow the live epoch, loss, validation F1, IoU and ETA metrics
- Run detection across the image set with "crackz detect run"; tiling at 512 pixels with overlap is applied automatically
- Review results in the galleries and summary tabs, where defect-detected cards are tinted so flagged findings stand out
- Export annotated PNGs and GeoJSON to your GIS pipeline, and generate the Word or PDF report for the audit trail
Pros & Cons
Pros
- Pixel-level, multi-class segmentation rather than a simple crack or no-crack verdict
- Trainable on the customer's own defect taxonomy, with a documented path from a 100-image pilot to production
- Can run on-premise and offline on operator hardware, which suits sensitive sites and locations without connectivity
- Source-available codebase, and the AI Pilot engagement hands over source code and trained weights
- Output lands directly in existing workflows: GeoJSON for GIS, plus Word and PDF reports for the audit trail
- The mask editor is claimed to be 50 to 80 percent faster than CVAT on dense crack imagery
- Reduces hazardous human intervention, with the publisher citing near-zero confined-space entries and no need to send divers or rope-access teams first
Cons
- No price is published for any product or engagement, so the entry ticket cannot be assessed
- No terms of service, privacy policy, legal notice or DPA is published at all
- No API and no API documentation, confirmed by probing both paths and subdomains
- No free plan and no free trial is announced, and there is no way to evaluate the tool without contacting sales
- Still at version 0.81, below a 1.0 release, despite being presented as in production
- The publisher is extremely small, reporting one employee and 563 thousand SEK of turnover in its latest published accounts
- Headline figures of 70 percent cost reduction, 5x faster cycles and 90 percent-plus accuracy come with no stated method, named dataset or third-party verification
Pricing & Plans
No price is published. There is no free plan and no free trial announced, and no pricing page exists — the "Pricing" link in the site footer points to the contact section rather than to a rate card. All three consulting engagement models carry the wording "Contact us for pricing", and Crackz itself is described only as a commercial source-available licence, with enterprise support and on-premise licensing handled through a sales conversation. The single monetary figure on the site, 25,000 to 75,000 USD, is a claimed annual saving for the customer's inspection programme and must not be read as a price.
- Crackz — no published plan or price
- commercial source-available licence
- with enterprise support and on-premise licensing arranged through sales
- AI Pilot — a two-week proof on the customer's data
- covering discovery and use-case scoping
- a working prototype
- quantified results for precision
- recall
- F1
- latency and cost
- a ship
- Production AI Sprint — an eight-week engagement adding production inference in the cloud or on-premise
- a retraining pipeline with evaluation gating
- an operator-facing interface or API
- documentation and handover
- and two weeks of post-launch support. Price on request
- Embedded AI Engineer — an ongoing retainer providing a dedicated senior ML engineer
- sprint planning and model reviews
- architecture office hours
- on-call cover for production incidents
- roadmap input and a quarterly review. Price on request
Data, GDPR & hosting
A consolidated view of how Crackz handles your data.
GDPR overview
There is no mention of the GDPR anywhere on the site, in either the English or the Swedish version. No privacy policy, no legal notice, no cookie information, no data protection officer and no Article 27 representative are published. One Swedish phrase can mislead an automated reading: Integritetskontroll appears on the Swedish homepage, but it refers to structural integrity inspection of chemical, refinery, wind and rail assets — not to privacy. The publisher is established in Sweden and therefore falls within the scope of the Regulation, but it neither claims nor denies compliance. The Content-Signal: ai-train=no directive in the robots.txt file governs reuse of the website's own content and says nothing about customer data.
Who owns the data?
The publisher puts no contractual document online at all — no terms of service, no privacy policy, no legal notice and no data processing agreement. Their absence is established rather than assumed: the sitemap lists only four URLs, and eighteen probed legal paths all return the site's catch-all page. Consequently nothing is stated anywhere about who owns imagery, models or detection output, what the publisher may do with customer data, or which third parties could receive it. The only adjacent fact is technical: Crackz can be deployed on-premise, and the companion Survey Ops product is described as running entirely on operator hardware. Anyone contracting for this software should obtain written terms directly, since none are published.
Reuse rights
Nothing is published on this point. With no terms of service and no privacy policy anywhere on the site, the publisher never states whether a customer may reuse imagery, trained weights or detection results without permission. Two facts point in the customer's favour but stop short of a licence: the AI Pilot engagement promises "source code and trained weights, yours to keep", and Crackz is described as a commercial source-available codebase. Neither is a published licence text, and the source-available terms themselves are not on the site. Note also that Crackz is a registered trademark of the publisher. Reuse rights therefore have to be negotiated and put in writing before any engagement.
Data retention & training
Hosting summary
No hosting jurisdiction is disclosed. The publisher names no country and no region where customer data would be stored, and publishes no privacy policy or subprocessor list that could establish it. What is documented is the deployment model rather than the hosting location: Crackz is offered on-premise or in the cloud, and the companion Survey Ops product is described as fully offline-capable, running on operator hardware with no cloud dependency. Azure Batch is mentioned, but as infrastructure the publisher uses for training work in its consulting practice, not as a stated hosting arrangement for Crackz customers. The marketing site itself sits behind Cloudflare and its mail is handled by Microsoft 365, which describes the publisher's own website rather than any customer data platform. Buyers with a data residency requirement will need to obtain it contractually, since nothing is published.
Things to keep in mind
Risks and trade-offs to weigh before adopting Crackz.
- The company identity on the site does not match the Swedish register. The site trades as "Anaxiatech AB" and publishes company number 556673-0056, but that number is registered to Pensionat & Konferenshotell Ullevi AB in Gamleby, with "Anaxia Tech" recorded only as a secondary trade name for IT consulting. No company named Anaxiatech AB exists in the register
- The site states it was founded in 2022, but the legal entity behind that company number was registered on 22 December 2004. The 2022 date matches new articles of association and the domain registration, that is the start of the IT activity, not the incorporation
- No terms of service, privacy policy, legal notice or data processing agreement is published, so a user has no written statement on data ownership, retention, sharing or deletion before committing
- Performance claims of 70 percent cost reduction, 5x faster cycles, 90 percent-plus accuracy and 25,000 to 75,000 USD of annual savings carry no stated method, named dataset or independent verification
- Automated defect detection on safety-critical structures invites over-reliance. A 90 percent-plus accuracy figure on validation data still means missed defects, and the software should complement rather than replace qualified engineering judgement and physical inspection
- The client logos shown, including SAAB, Ericsson, Telia, Vodafone and Discovery+ / Warner, are presented as the publisher's past and current consulting references, not as Crackz customers
- The publisher is a one-person operation at version 0.81 of its flagship product, which carries real continuity and support risk for a long-lived inspection programme
Setup & Integrations
Technical difficulty
Technically demanding. There is no self-service sign-up, so setup begins with a sales conversation. The software is then installed on the customer's own hardware as a desktop application, a command-line tool or a Docker container, ideally with a CUDA or ROCm GPU, though CPU execution is supported. Projects are configured through a YAML file. Using a pre-trained checkpoint is the fast path, described as production-ready in days with no training data required. A custom model is considerably heavier, requiring 100 to 500 or more manually annotated images. A technical team is assumed throughout.
Deployment
Behind Crackz
Resources
All the official URLs gathered for verification and reference.
Alternatives
Tools that compete with or complement Crackz.
Frequently asked questions
What defects can Crackz actually detect?
What surfaces and industries is it used on?
Do I have to train a model myself?
How is the software delivered?
Can it run without an internet connection?
What output formats does it produce?
How much does Crackz cost?
Is there an API?
Is there a mobile app?
Who publishes Crackz?
Should you pick Crackz?
Crackz is a narrow, serious tool rather than a general-purpose assistant, and it is best judged on that basis. The engineering is specific and credible: pixel-level segmentation on a documented SegFormer and PyTorch Lightning stack, a genuine annotation editor, trainable multi-class taxonomies, and output that lands where inspection work actually happens — GeoJSON for GIS pipelines and Word or PDF reports for the audit trail. The ability to run on-premise and offline is a real differentiator for defence, utility and port operators who cannot send imagery to a cloud service.
Against that stand two substantial reservations. The first is commercial opacity: nothing is priced, there is no free plan or trial, and no self-service path exists, so any evaluation begins with a sales conversation. The second is more serious for a buyer's legal team. The publisher puts no terms of service, privacy policy or data processing agreement online, and the company identity displayed on the site does not match the Swedish register: the company number shown belongs to an entity registered under a different name, with "Anaxia Tech" recorded only as a secondary trade name for its IT-consulting activity. Headline performance figures are also unsourced.
The result is a product whose technical proposition is well argued but whose contractual and corporate framing is not documented at all. For an inspection team with in-house technical capability and a procurement process able to demand written terms, Crackz is worth a conversation — particularly for the harbour and bridge use case it was built for. For anyone needing to assess cost, licensing or data handling before making contact, the site simply does not provide enough to go on.
- Choosing a selection results in a full page refresh.
- Opens in a new window.