The research operation
An AI research lab is a strange thing to run from Anchorage. It is also the entire reason we can promise what we promise.
Northtek Labs is the research half of the company. It exists because the things our clients need most - an agent whose actions can be audited, memory that never leaves their building, output that can be checked - were not available as products when we needed them. So we built them, and we published them.
Written by Kristian Baer, Northtek · Anchorage, Alaska · Updated 2026-08-21
The short answer
What is Northtek Labs?
The research arm of Northtek, based in Anchorage, which builds and open-sources AI infrastructure. Four things run there. Kryos is a programming language and compiler built so an agent’s actions can be audited line by line. GENOME is an offline memory server, benchmarked in public against a commercial competitor on standard datasets, with the harness released alongside the numbers. FACTGATE is a verification gate that checks model output against its sources, and we publish its false-accept rate including the cases it misses. aiproof is a prompt-injection scanner, written because we deploy agents that read untrusted documents and email. There is also a patent filing on the memory architecture. All of it is readable before anyone talks to us, which is the point.
Why a lab at all
Every piece of this was built because a client needed it and it did not exist.
This is not research for its own sake, and it is not a marketing exercise. Each project traces back to a specific promise we could not otherwise keep.
A Native corporation asks where shareholder records physically live. The honest answer for a firm renting a third-party memory layer is "wherever that vendor decides", so we wrote our own with an offline mode. A federal contractor asks how they can be sure a proposal draft did not invent past performance, which is a False Claims Act question rather than a quality one, so we built a gate that requires every claim to trace to a source document and published how often it fails. An auditor asks how an agent reached a conclusion, so we built a language where the answer is a readable sequence of steps.
The pattern holds throughout: the research exists because a regulated client asked a question that products could not answer.
What publishing costs
We publish the results we lose on, which is the part that makes the rest believable.
Our GENOME write-up reports where it beats an established commercial competitor on ingestion and where it only reaches parity on accuracy. FACTGATE ships with a false-accept rate rather than a claim of correctness. Those numbers are less flattering than the alternative, and they are the reason the favourable ones are worth anything.
A benchmark published only when it wins is a selected figure. A benchmark with no harness is a marketing figure. We release the harness so anyone can run it, disagree, and say so - which happens, and is useful.
What it means for a client
You will never use any of this directly. It changes what we can offer you.
No client engagement ships a research project, and we would push back on anyone who suggested it. What the lab changes is the shape of what we can promise: a workload that runs entirely inside your own boundary is a configuration choice rather than a custom project with a custom price, because we own the memory layer instead of renting it. An audit trail is a property of the system rather than a line in a proposal. An accuracy figure comes from a harness rather than an impression.
It also changes what happens when your problem does not match an existing product, which in Alaska is often. A firm that assembles tools is bounded by what those tools expose. A firm that writes them is not.
Apply it
What is public, right now
Open any of these before you contact us. That is what they are for.
-
01
Kryos
A programming language and compiler for auditable agents, with its full commit history. Writing a language is difficult to fake, which is most of why it is worth pointing at.
-
02
GENOME
An offline memory server, benchmarked against a commercial competitor on public datasets, harness and complete results published including the ties.
-
03
FACTGATE
A verification gate for model output, published with its false-accept rate and worked examples of what it lets through.
-
04
aiproof and SOFAR
A prompt-injection scanner for agents that read untrusted input, and long-context memory research with the benchmark harness released alongside the numbers.
Straight answers
Is there another AI research lab in Alaska?+
The University of Alaska does genuine research across several departments, and it is a different kind of institution with different aims. Among firms selling AI services in this state, we are not aware of another with a public codebase, published benchmarks, or a patent filing on infrastructure they wrote. If that changes we will say so, and it would be good for the state if it did.
Does the lab work with the university or other researchers?+
We publish openly, respond to critiques, and have had our numbers argued with, which is the useful part. Formal collaboration is a conversation worth having and not something we are going to claim on a web page before it exists.
Can we license the lab technology directly?+
Most of it is open source, so you can already use it - that is what publishing it means. The patent filing covers the memory architecture specifically. If you want to build on any of it, the fastest route is to open the repository, and get in touch if you hit something.
How does a small Alaska company fund this?+
Client work funds it, which is the honest and unglamorous answer. It also means the research is disciplined by real deployments rather than by what is interesting - every project here started as something a paying client needed and could not buy.
Is any of this actually used in client work?+
Yes, and selectively. The memory layer and the verification gate are the two that appear most often, because data residency and citation are the two things regulated Alaska clients ask about first. We do not deploy something into your operation because we happen to have written it.
What are you working on now?+
The active thread is measurement rather than new infrastructure: the Alaska AI Operations Index, a survey producing the first real count of AI adoption in this state by industry, and a document difficulty study measuring extraction accuracy on the paperwork Alaska actually runs on. Both publish their methodology and raw data.