We always start in the same place: clear goals, measurable acceptance criteria, and a strategy that holds all the way to production. Then, and only then, do we build the agentic solution, model, and software around it.
Always built together with your team, integrated into your core infrastructure, and fully owned by you end to end.
We help you identify processes worth automating, select the right agentic framework and build robust AI agents that fit your operations. And just to make it clear – you own everything, models and code.
We’ve built entire production systems prompted through AI agents for some of Sweden’s largest companies. Including automated tests, documentation, and production-quality code delivered several times faster than conventional AI-assisted development.
We have deep expertise in model architectures and foundation model training. We design and train transformer models for generation, outlier detection, classification, analysis, and multi-task learning – enabling a single model to handle multiple objectives efficiently.
Selected clients
Partners
Network
AI agents do the heavy lifting and write the code, while we own the direction, the reviews, and the decisions. That is how we built everything from one of the world's most popular Obsidian plugins to automating complex production systems for our clients.
A 64-minute YouTube movie was dubbed into 20 languages with Johan's cloned voice, in sync with the subtitles, in two days. Here is the full pipeline — and everything that went wrong along the way. The subtitles that broke in every language, the sleepy voice, the word that meant both tag and day, and how four audit agents found the flaws across twenty languages.
The first issue of Tech Insights has gone out through our new engine. Hand-written markdown in, one immutable API request out to over 3,000 subscribers. This is the story of how we built the system: the consent database, the armed release, the Monday 06:00 machine, and how we closed the windows that threatened to send the wrong thing.
Agentic AI is when models plan, call up tools and applications, and take actions inside a defined workflow. We deliver with integration, evaluation, guardrails, and a design that holds up in production.
We prioritise work that’s possible to deliver and operate – usually repeatable processes with clear inputs/outputs, testable quality, and realistic integration points. We prioritise together with your executive team, and every delivery is built against explicit acceptance criteria before we scale.
Not always. If an off-the-shelf, general-purpose LLM meets your quality, privacy and cost constraints, we’ll use it. We build custom models when you need stronger performance, tighter controllability, or have ownership and deployment requirements that demand it.
You do, always. We deliver so you can run, evolve and extend the system, including models, code and documentation.
We treat them like production systems. That includes evaluation, test suites, observability, clear boundaries for what the agent may (and may not) do, and safe fallbacks when uncertainty is high.
We design for your environment and requirements, from data handling to auditability and deployment patterns.
Because an AI transformation cuts across the whole organisation, we start with your executive team: together we identify processes, candidates for agentic AI, and the strategy for your software development. When we know what will make a difference, the work starts on-site with you, in joint work streams. The first delivery is typically in production within three to six months, then the teams continue with the next delivery in the queue. The knowledge stays with you, and so does the result.
Short takes on what matters in the world of AI. What’s noise, and what’s next – written weekly by our CEO Johan Sanneblad.
OpenAI previews an Ultrafast tier running GPT-5.6 Sol up to 750 tokens per second; Anthropic adds invisible watermarks to Claude and makes auto mode default; Lovable raises 400 million dollars at a 13.3 billion valuation.
Read moreGoogle moves Demis Hassabis to Alphabet Chief Scientist and reshuffles DeepMind. OpenAI updates GPT-5.6 Sol and offers unlimited free Luna chats. The EU AI Act enforcement begins, and Stanford uses AI to design 16 working viruses.
Read moreOpenAI says its Astra model solved ten open math problems and GPT-5.6 Sol cut serving costs 20 percent. Anthropic reveals Claude cyberattacks on real companies, and Moonshot AI releases Kimi K3 weights.
Read moreOur symbol is inspired by the “Einstein tile” – a single shape that can tile the plane without repeating. It’s a reminder that real capability often starts simple with one robust building block, assembled into patterns that scale beyond templates.
