We build Crackz — AI defect-detection software in production across harbors, bridges, factories and energy assets. We also take on senior consulting engagements across AI, system architecture, DevSecOps, simulation, defense, and autonomous systems — from the same Stockholm studio, led by a systems architect with 40+ years of experience.
AI-powered defect detection — cracks, corrosion, surface flaws, weld defects. Built for harbor and bridge infrastructure; deployed across factories, energy assets and manufacturing QA. Desktop, web, and CLI from one source-available codebase.
Process thousands of inspection images in hours instead of weeks. Detect issues before they become critical. Document everything for the audit trail.
Single pane for project state — model config, training schedule, detection parameters and data paths. Everything reproducible from one YAML.
Brush-driven annotation built for production scale. 50–80% faster than CVAT on dense crack imagery, with mask propagation across image series.
Live training with live metrics — Epoch · Loss · Val F1 · IoU · ETA. Built on PyTorch Lightning; checkpoints record their annotation schema so incompatible models are flagged automatically.
Galleries, summary tabs, severity buckets. Defect-detected cards are tinted in a dark green so flagged findings read at a glance across thousands of images — whether they're harbor pier scans or factory weld inspections.
Drop in infrastructure, manufacturing or facility imagery. Automatic raw/train/val/test split, EXIF preserved end-to-end.
Mask editor with brush, propagation across image series, and category-aware multi-class taxonomies.
SegFormer-B2 by default; swap encoders, override hyperparameters at runtime. PyTorch Lightning underneath.
Run across thousands of images. Severity buckets, annotated PNGs, GeoJSON output ready for QGIS or GIS pipelines.
Quick deployment with ready-made checkpoints. Suited to standard crack-detection scenarios across concrete and steel.
Maximum accuracy for your domain — bespoke imagery, unusual defect types, multi-class taxonomies.
Every capture, from field to archive — catalogued, geotagged, and ready for analysis. The data backbone of the AI Slam platform, built for subsea expeditions.
Resumable field uploads over Starlink. Operator inbox approval. USBL geotagging. Geospatial map of every capture. Natural-language search with cited answers.
Browser or CLI upload from vessel or shore. Resumable over Starlink — a dropped connection never loses progress.
Packages stage in an inbox. AI pre-fills metadata from the file listing; an operator approves before anything reaches canonical storage.
USBL track data writes real GPS positions onto images. Every capture indexed by project, site, date, and geolocation — visible on a live map.
Ask questions in natural language. Answers are grounded in the archive with source citations. Hand structured captures to Crackz for defect detection.
We're a small Stockholm team building production AI for defect detection — across infrastructure inspection, manufacturing QA, and energy assets. Crackz is the flagship; the capabilities behind it are available for engagement.
Drone and handheld imagery into measured, ranked defect patterns — without sending divers or rope-access teams in first. The use case Crackz was originally built for.
Scratches, voids, weld defects, corrosion, paint failures, casting porosity. Trainable on your taxonomy — multi-class segmentation, not just crack/no-crack. Inline or post-process.
Remote analysis of imagery from maintenance teams, robotic inspectors, or autonomous drones. Predictive maintenance, condition-based renewal — instead of emergency response.
Crackz is the commercial product. These are the research and tooling efforts that feed back into it — marine archaeology, 3D mesh tooling, and other custom systems.
Commercial source-available AI defect-detection software — infrastructure and manufacturing. Desktop · web · CLI · Docker.
Long-running partnership with the two sister companies behind the M/S Estonia reinvestigation, the Kraveln (1524) scan, and BBC / Discovery deep-water assignments. We run their IT, camera, networking and 3D-reconstruction pipelines.
Local-first catalog and ingest platform for marine survey data — moves field captures from inbox to canonical project storage with geotagging, automated processing pipelines and a map-based operator dashboard. Built for subsea expeditions, feeding into the wider Crackz program.
An open, vendor-neutral topside operator platform for ROVs — one cockpit for the vehicle, cameras, sensors and positioning. A time-synchronized device platform turns a pile of independent hardware into a single, hot-pluggable console, with navigation warnings, computer vision, mission planning and photogrammetry layered on top. First target: the Ocean Modules V8 over its raw link.
Unreal Engine ROV training environment for subsea operations. Photogrammetry-captured wreck sites, physics-accurate underwater dynamics, turbidity modelling — mission rehearsal without the boat.
Terminal monitor and APRS / GPS bridge for Meshtastic LoRa mesh networks. Off-grid comms tooling for maritime tracking, drone telemetry bridging, and HAM-bridged field operations.
Anaxiatech AB is a Swedish technology company — founded and led by Theresia Lundgren, a systems architect with 40+ years of experience across AI, DevSecOps and applied R&D. We design and ship production software for problems where the wrong answer is expensive: infrastructure inspection, manufacturing QA, defense systems, marine archaeology.
Hands-on work, end to end: we write the model code, ship the desktop app, run the Azure Batch jobs, and sit with operators while they review the results. Past and current clients include SAAB, Ericsson, Discovery+ / Warner Brothers, and Swedish government agencies. Long-running technical partnership with Ocean Discovery AB and SubBaltica AB on subsea expeditions.
We're the team that built Crackz. We take on consulting engagements across AI, system architecture, DevSecOps, simulation & digital environments, government & defense, and autonomous systems — for organisations that need a serious answer to a serious question. No staff augmentation. No PowerPoint. Delivered.
We've shipped real-world segmentation pipelines processing thousands of inspection and QA images per run. We know where the failure modes are: dataset shift, label noise, overconfident models on rare classes, ONNX export gotchas, why your A100 is sitting idle.
Defect detection, crack and corrosion segmentation, weld and casting inspection, surface-finish QA. The vertical we know cold — Crackz is the public version. We can build the private version for your domain in 8–12 weeks.
Where LLMs actually earn their keep: structured extraction from technical reports, RAG over engineering knowledge bases, agentic workflows for code review and triage. We pick the model that fits, not the one in the press release.
Training pipelines that retrain on schedule, evaluation that catches regressions before users do, deployment that survives the model getting bigger every six months. Azure Batch, Lightning Studios, self-hosted runners for sensitive data.
Should you fine-tune, RAG, prompt, or just buy the API? Two-day discovery engagements for technical leaders who need a defensible answer for the board. We map the option space, cost each path, and write up the recommendation.
For domains where you can't easily collect a thousand labelled images — underwater, defense, hazardous, pre-production manufacturing — we generate it. Photogrammetry-driven training data, ROV simulation, reinforcement-learning environments.
Fixed-scope for the first phase. Continue on a retainer if it's working. We don't do staff augmentation, hourly billing, or open-ended scopes.
A focused two-week proof on your data. We pick one ML question, we answer it with numbers.
Contact us for pricing
An eight-week engagement to ship a vertical slice of production ML — model, infra, and a UI to drive it.
Contact us for pricing
Ongoing senior ML capacity for teams that need depth without hiring a new headcount.
Contact us for pricing
Pricing, proof-of-concept projects, enterprise support, and on-prem licensing. We respond within one business day.