Build intelligence you can own.
Proprietary AI systems for the documents your business runs on. Designed around your data, your evaluations, and your freedom to change models.
- 01 · Production lineage
- Multi-year aviation system
- 02 · System posture
- Model-agnostic by default
- 03 · Operating region
- Nordic · EU-ready
- 04 · Starting point
- Architecture review
Products / 03
Find it. See it. Set it in motion.
Three focused products for knowledge retrieval, visual search, and transparent agent execution.
Explore all productsThe operating system
A system, not a stack of subscriptions.
The token factory is a thin control layer between your documents and the model frontier. You own the parts that create durable advantage; providers stay replaceable.
- Replaceable models
- Replayable decisions
- Traceable document lineage
- Explicit human escalation
- 01
Context engine
ingest → embed → retrieve
Your corpus enters with provenance, permissions, and an owned schema. Nothing useful starts with an ungoverned vector dump.
- 02
Prompt compiler
intent → template → tokens
Business intent is separated from provider syntax, so the same task can move between model families without a rewrite.
- 03
Model router
route(task, cost, SLA)
Each task goes to the model that fits its quality, latency, residency, and cost constraints—not the vendor you picked last year.
- 04
Eval harness
replay → score → gate
Golden cases make every change measurable. Model upgrades become controlled releases instead of hopeful API swaps.
Selected work
High-stakes documents. Concrete systems.
The same owned architecture takes different shapes. The line is kept clear between work delivered in production and a strong expert approach.
View all workField notes
The thinking ships too.
Notes from building applied AI: architecture, economics, regulation, and what survives contact with production.
All field notes“The methodology comes from building a document intelligence system that had to survive production—not a pitch deck.”
Blue Abraham is built by Martin H. Normark. The token factory architecture grew out of years spent building AI for aircraft technical documentation at a Nordic aviation company, later acquired.
That work crossed model generations, document formats, and operating constraints. The parts presented here are the parts that held up.
- Built in
- Production
- Domain
- Aviation
- Delivery
- Founder-led
Ways to work together
Engage at the point of highest leverage.
Most work begins with an architecture review: short enough to be decisive, deep enough to change the build.
Services and pricingA useful first move
Bring one difficult document problem.
An architecture review takes one to three weeks, ends in a written decision document, and creates no obligation to keep working together.
Start the conversation