AitiumResearch
Useful AI that runs on hardware you control, with limits you can enforce and results you can check. Aitium builds small specialized models, the engine that serves them on one or two 24GB GPUs, and the runtime that keeps AI agents inside the boundaries their operators set. We publish what we measure, including what fails.
Research areas
Can a small model, sized for one local GPU, do a narrow expert task well enough to trust?
CyberGuard targets security triage and threat analysis, measured on precision, recall and false-positive rate before any release.
See the workCan matching DNA variants to medicine compatibility become a model query without losing expert-level accuracy?
Cogenics validates variant-to-drug associations before any model is trained, using CPIC prescribing guidelines as the evaluation axis and ClinVar as the reference dataset.
See the workHow capable a model can run privately on one or two 24GB consumer GPUs without silently degrading precision?
Ember is a registry-driven gateway and GPU worker, live on Aitium hardware and serving an OpenAI-compatible API.
See the workHow do you give an AI agent real autonomy while keeping its reach, spend and actions enforceably bounded?
Foxtail Harness puts a tool registry, delegation scopes, atomic budgets, review gates and an audit trail between agents and production systems.
See the workRoadmap
Funding focus
Outside funding goes to three things, each tied to a public result on our roadmap.
GPU time to train CyberGuard and run it against Cybench, CyberSecEval and the AWS Deception Benchmark.
Result Full evaluation results published, misses included.
Q4 2026
Blinded review by pharmacogenomics specialists, checking variant-to-drug associations against CPIC guidelines before any model is trained.
Result Validation published, with a go/no-go training decision.
Q1 2027
Engineering time to harden the governed agent runtime — registry, budgets, review gates and audit trail — for outside teams.
Result Early access for external teams.
Q4 2026
Evidence
We publish what we measure, including the misses.