LANARS
Firmware and software
- Firmware for the floating unit
- Embedded vision stack
- Annotation platform (web app + APIs)
- Cloud data path
- Operator-facing tooling
- OTA pipeline
Case study · LANARS × SelectAI
AquacultureComputer visionEdge AI
Six months from concept to floating hardware that meets Mattilsynet's clinical taxonomy .
LANARS built the firmware and the software for SelectAI's in-cage CV unit, in partnership with AI Experts. Live pilots on Norwegian salmon farms today; backed by Innovation Norway; targeting regulator-grade automated lice counting by end-2026.
At a glance
- 6 monthsconcept to working hardware and validated software
- 15,000+fish analysed and archived
- ~1,000 vs 20fish per session vs manual netting standard
- 3 camerassynchronised, multispectral, tri-angle
- 5 classesMattilsynet lice taxonomy, no internal mapping
- 2026target for Mattilsynet automated-counter approval
Context
Norwegian aquaculture needed a harder tool
Sea lice cost Norwegian salmon farming billions every year. A single adult female louse can produce up to fifty offspring per day in summer, and the manual standard — netting and counting fifteen to twenty fish per cage every week — samples too few fish to catch outbreaks before they bloom.
Weekly reporting to Mattilsynet is statutory. The Norwegian Directorate of Fisheries has publicly stated that count accuracy across the industry is too low and that tougher requirements are coming.
The opportunity wasn't another dashboard. It was a measurement tool precise enough to act as the regulator-grade baseline.
Three teams, one stack
Who owned what
AI Experts
Models and validation
- Model architecture and training
- Taxonomy classifiers
- Welfare-indicator detectors
- Validation methodology
SelectAI in-house
Product and operations
- Product vision
- Aquaculture partnerships
- Hardware fabrication and assembly
- Regulatory engagement
- Field operations
How it works
How a fish becomes a report
Patent-pending net guidance feeds salmon into a lit analysis channel one at a time. Three synchronised cameras image left, right and underside in a single capture. The annotation pipeline maps detections directly to the five Mattilsynet lice classes and the Laksevel welfare scale — no internal mapping layer.
- 01
Guided channel
Patent-pending net system; fish swim through voluntarily, no anaesthesia, no manual handling
- 02
Tri-camera capture
Three synchronised lenses, multispectral lighting, motion-triggered frames
- 03
Vision engine
On-device pre-processing, cloud-side inference, biologist-reviewed annotation loop
- 04
Mattilsynet-ready report
Weekly lice-per-fish, Laksevel welfare summary, image-traceable assessments
Inside the unit
Engineering decisions, on the record
- 01
Tri-angle synchronised cameras
One fish, three angles, logged once. Single-angle competitors double-count.
Three lenses fire in the same frame window
- 02
Multispectral lighting
Standardised illumination plus cameras registering up to 60% more colour shades than the human eye.
Picks up signals an inspector would miss
- 03
Motion-triggered capture
Compute and storage are spent only on frames with a fish in them.
No background imagery in the pipeline
- 04
Edge-first operation
No power or internet on the cage. The unit runs off-grid and syncs on retrieval.
Built for tough maritime conditions
- 05
Salt-water hardening
Tested through a Karmøy winter. No drama on the hardware side.
Validated on Varde Fiskeoppdrett, December 2025
- 06
OTA-deliverable capabilities
Disease detection, biomass, respiration, other species — roadmap features ship as software updates to the same unit family.
Same hardware, new abilities
The foundation
Why image quality is the foundation of reliable AI in aquaculture
Artificial intelligence is only as good as the data it learns from. When training a model to detect sea lice or welfare injuries, the process relies entirely on labelled images — photographs where a human expert has already identified and marked what the AI is expected to learn to recognise. If those images are blurry, poorly lit, or only show one side of the fish, the labels are uncertain at best and wrong at worst.
Before asking what an AI can detect, ask a simpler question: can a trained expert clearly see and identify the lice or the injury in this image? If the answer is no, the AI has nothing reliable to learn from — and nothing reliable to deliver. This is why image acquisition matters as much as the algorithm itself.
SelectAI Focus is built around this principle. As each fish passes through the channel, three synchronised cameras capture the entire surface under standardised lighting — dorsal, ventral and lateral, every fish, every pass. The results are based on full statistical coverage of the population, not on predictions extrapolated from a partial view. For sea lice counting and welfare assessment, that's the difference between knowing and estimating.
Consistent, high-quality images collected under identical conditions over time do more than support today's reporting requirements. They build a growing, labelled dataset that makes new models possible — early disease detection, biomass estimation, respiration monitoring. The value of an aquaculture AI platform is not only what it does today; it's the foundation it builds for tomorrow.
Annotation suite
Built for biologists, audited for regulators
The annotation tool is where biologists turn captured frames into labelled data. Every fish is reviewed across three synchronised camera angles, lice are categorised against Mattilsynet's clinical taxonomy, and welfare indicators are localised pixel by pixel — with timestamp, operator, GPS and source frame travelling with every record.
Tri-angle pane, one specimen
Left, right and underside annotated together so the same fish is never double-counted.
Clinical lice taxonomy
Stuck · Movable · Mature female · Skottelus · Free-swimming. No internal mapping layer.
Welfare-indicator boxes
Wounds, fin damage and scale loss localised pixel-accurate for downstream training.
Audit-grade reports
Every annotation carries timestamp, operator, GPS and source frame.
Measured, not promised
Outcomes
- 1,000 → 20fish per session vs manual standard
- 15,000+fish archived, growing weekly
- 100%surface coverage per pass — statistical, not extrapolated
- Days earlieroutbreak warning vs manual counts
- 2 product linesshipped on the same stack: Focus and Multi Grader
- End of 2026headline goal: Mattilsynet approval
From the founder
"The aquaculture industry faces some of the toughest ecological challenges of our time — yet most farms still rely on tools and methods that haven't meaningfully changed in over a decade. We're not here to sell another promise of a digital future. We're here to deliver a real result today: help the farmer act before an outbreak, lift fish welfare in measurable ways, and do it at a price that actually makes sense for the industry."
Tech stack
- Computer vision
- Edge AI
- Embedded firmware
- Multispectral imaging
- Annotation pipeline
- OTA
- Cloud data path
- React / Next.js
- TypeScript
- Flutter web
Looking to ship something this hard?
We do firmware, embedded vision, AI integrations and the platforms that make them usable. Tell us what you're building and we'll come back within a business day.
Discover more cases
