NORTHTEK Labs.
Open-source tools, AI research, and internal infrastructure that we use to ship faster and build smarter for our clients.
RAIN, exact reasoning without neural networks
An open-source reasoning system that solves three published compositional-generalization benchmarks at or near 100%, in seconds on a plain CPU, with no neural network doing the reasoning. Frontier AI models measured on the same held-out tests score 36% to 86%. Every number reproduces from a fresh clone.
RCK, the Resonant Cognitive Kernel
A reasoning engine that answers from stored facts instead of guessing, shows the chain of facts behind every answer, and says "I don't know" when it doesn't. It runs on a plain CPU, learns a new fact instantly without retraining, and holds 100,000 real-world facts in 535 MB with sub-millisecond lookups.
SOFAR, so a transformer stops forgetting
A drop-in adapter for any HuggingFace transformer that helps it find facts buried deep in a long prompt instead of losing them. On the published GPT-2 tests it lifts needle-in-a-haystack recall by a third to double, while adding about three hundredths of one percent to the model's size.
Kryos, a language that has to ask permission
A compiled programming language where code cannot touch the network, the disk, or your data unless it declared that intent up front, and the compiler refuses to build it otherwise. Built for the moment when AI agents run real actions, it is the difference between trusting a promise and enforcing a rule.
Lemmas, reliability primitives for any LLM
Small, dependency-light modules that make any language model API more trustworthy: verifying its own answers, checking for consistency, and catching quality drift. No framework to adopt and no lock-in, they wrap whatever provider you already use.
FACTGATE, the AI fact-checker that is not allowed to guess
An open-source gate that checks an AI assistant's answers against the facts of your document and returns one of three verdicts: verified, blocked, or held for a human. No judge model, no similarity thresholds, no guessing. Across 110 hand-reviewed answers on four real documents, it never once confirmed a wrong answer.
The Alaska Lead Engine
A free 12-page field guide to the five-part build that turns a local business into the obvious choice in its market - getting found, converting, content, paid ads, and follow-up. Every part ends with a checklist you can run this week.
More open source
Tools we built for our own work and released for anyone to use. Every one is public on GitHub.
Ghost
RustThe computer-use layer for AI agents on Windows and Linux. It operates any app, even ones with no API, in the background without taking your screen or cursor, and proves every action actually happened.
Evolve
RustDrop-in A/B evolution for AI coding assistants. Keep using Claude Code, Cursor, or Aider exactly as you do, and it quietly learns which setup works best for your codebase, then promotes the winner. No cloud, one local file.
Praxis
TypeScriptA memory for the things an AI assistant has proven it can do. The twist most memory tools miss: it only saves a skill after the assistant has actually run it and shown it works. If it does not pass the check, it does not get saved.
aiproof
MakefileESLint for AI prompts. A static analyzer for the prompts you feed to language models, it finds real bugs before you spend a single token.
Lossless Context
TypeScriptA context ledger for agent file reads with cryptographically signed receipts of exactly which file versions the model was shown. Refreshingly honest about its own benchmarks, including where it does not help.
Claude Code Flywheel
JavaScriptMeasure which Claude Code skills and agents actually earn their cost. A small, fully local observability layer for your AI setup, with zero runtime dependencies.
Aurora3D
TypeScriptA react-three-fiber library and Claude skill for generating Spline-quality 3D web visuals in code, the kind of motion we use on client sites.
OVERSTORY
TypeScriptA knowledge tree of your codebase where every claim carries a verifiable receipt: it hashes the exact source lines each claim cites, so stale docs fail loudly instead of lingering. Local-first CLI plus an MCP server for agents, on npm as @northtek/overstory.
GENOME
PythonAn auditable memory layer for AI agents that ingests locally with zero LLM calls, matches Mem0 on benchmark accuracy at a fraction of the ingest cost, and can answer what it believed at any point in time. On PyPI as genome-memory, MCP server included.