Your inspection camera lets defects through, or throws good parts away.

You already have a smart-camera inspection running on the line, and it is not good enough. The usual cases:

• Too many escapes: defects reach your customer.
• Too many false rejects: good parts go to scrap or rework.
• A new defect class that the current model has never seen.
• A model that works on a GPU and not on the camera.

The deep-learning tools that come with the cameras handle the easy cases well by now. Volary takes the remainder: the cases where the tool cannot reach your escape or false-reject target.

Volary AB is a one-person company. The work is done by Dominic Portain, who spent five years building inspection AI inside a camera maker and a scanner maker before starting the company.

SICK IVP

On-device detection for the Inspector line: rugged smart cameras that detect, count and check small objects above conveyor belts. I worked on the training pipeline (including quantization-aware training), helped build the end-to-end detection benchmark suite, and designed datasets with the single purpose of bringing the detection system to its limits. They found two unhandled edge cases. One of them was fixed with a rotation augmentation that I implemented in the camera firmware.

Woodeye (MiCROTEC)

Line-scan scanners that grade wooden boards at high speed. At a career fair I asked one of their vision engineers which problem they could not solve. The answer was light knots in oak, which conventional image processing could not catch. I got a fixed-term contract, showed a working object detector within six months, and was hired permanently.

After that the detector went into production and grew to multiple defect classes plus a classification network. I ran the labeling operation (hundreds of physical boards, professional labelers), built the training hardware, and grew the AI team from one to five.

Volary also builds apps for iPhone and Apple Watch, in Swift and SwiftUI. Inklight is on the App Store; the others are work in progress. Each one has a page that says what it does and what works today.

• 

  Inklight

  Reads your own EPUB, PDF, text and Markdown files aloud with a neural voice that runs entirely on the iPhone, offline. Free for 10 minutes a day, with paid plans for more. On the App Store.

• 

  Tolk

  Listens to Swedish spoken in the room and shows it in English, sentence by sentence. Early build.

• 

  StoryReader

  A reader for plain-text stories in an iCloud folder, with a rating at the end of every story. First milestone built.

• 

  TimeBucketDrip

  A weekly time budget that treats hours like money. Runs as a web app, the iPhone version is not finished yet.

• 

  RandomWords

  One button that shows random words, to bring back a dream or another weak memory. The first version works, the next one is specified.

More about the iOS work

Volary AB is registered for consulting in artificial intelligence and software development. Inspection is the specialty. This is the rest of what the company does.

AI workflows and agents

Building, deploying and maintaining LLM-based workflows and agent systems for small and medium-sized companies, either directly or as a subcontractor to an agency. Volary worked this way for NineTwoThree, a US AI agency: embedded with the agency's clients, with projects owned end to end (discovery, product design, build, delivery, client communication). The details of that work are confidential.

What I bring to this kind of work: document extraction where every extracted number links back to the page it came from, evaluation harnesses with blind judges and leak checks, multi-agent workflows, retrieval, and local inference.

Software and app development

Custom software around data and models: pipelines, internal tools, dashboards and test suites. Mostly Python, plus TypeScript, Swift, Lua and C++ where the target needs it.

App development for iPhone and Apple Watch is part of this. Volary's own apps are described under iOS apps.

Embedded and electronics

The registered scope also covers the design and manufacture of electronic products with built-in AI. Nothing is on sale yet. The background for it exists, though: firmware work in Lua and C++ on industrial cameras, and small-batch electronics assembly (pick-and-place, reflow, microsoldering) at an audio hardware startup.

Dominic Portain is the sole owner and the only employee of Volary AB. I am an applied AI engineer with a statistical core: my MSc in Cognitive Science was mostly applied statistics, signal processing and programming.

My jobs so far follow one pattern. I enter a domain that is new to me, learn its logic, build a working AI or data system for a real problem, and explain it to the people who have to trust the output. That has been neuroscience, energy trading, wood defect detection, industrial smart cameras and LLM products.

The statistics are the reason my test sets measure rates per defect class. I look at residuals, question distributional assumptions, and go looking for the place where a model breaks.

I work in English and German. More detail is on LinkedIn.