A private AI core for reliable software agents.
NURM runs, verifies, measures and improves software agents — built to run where your code lives. Your code never leaves your infrastructure, and no claim ever outruns its evidence.
Obscyra Technologies L.L.C-FZ · Dubai, United Arab Emirates
Not a roadmap. A working factory.
Everything below is running now, on our own infrastructure.
Sovereign, end to end
Our own fine-tuned model, our own training stack, our own sandboxed executor, our own grading. No external AI APIs in the runtime loop — frontier models are used only as offline teachers during development, never inside the product.
model · training · execution · evidenceA self-improvement loop that has already closed
The factory mined its own model's failures, retrained it on its own verified traces — and the promoted model solved 17/17 held-out failure tasks its base model solved 0/17, with a clean regression gate across eight capability facets.
internal gold-free held-out · regression-cleanHonesty enforced by code, not policy
Runs are pre-registered before they start. Solutions that cheat — network escapes, test tampering — are detected and discarded, not rewarded. Promotion gates have blocked our own weak results; recorded negatives are part of the method.
pre-registered · anti-cheat · gated promotionRadical cost asymmetry
Prototype-stage spend, by design: the entire training-lab history to date ran on ≈ $250 of rented GPU time — one founder, an AI agent workforce. The point is the slope: a verified improvement cycle costs dollars, not teams.
solo founder · agent workforceRun → Verify → Measure → Improve
Run
Agents work in controlled, disposable sandboxes — with the tools they are allowed, and nothing else.
Verify
Outcomes are graded against executable evidence — tests, checks, artifacts — never against model opinion.
Measure
Every capability claim is a paired delta: same tasks, same budget, base versus base+NURM.
Improve
Verified failures become training fuel. What fails today is retrained to be solved tomorrow.
One number, honestly produced.
Most AI claims are demos. Ours is a measurement — and until it lands, this slot stays honest about that.
Our first externally checkable paired result is being produced now: a frozen task set, identical budgets on both arms, executable-test grading, pre-registered before the run. It will be published here with its full protocol — including the runs that failed.
The delta ladder — every claim names its rung
- D1–D2 · Runtime & skills liftin measurement
- D3–D4 · Trained-weights liftinstrumented
- D5 · Full stack, end to endthe headline
Leverage without surrender.
Engineering and AI-first teams — roughly 20 to 100 engineers — that want real leverage from agents without giving up control. Especially teams whose code cannot leave their environment: regulated, privacy-constrained, or simply sovereignty-minded.
Bring a private repo. NURM runs where your code lives.
Built solo, with an agent workforce.
Planning, training, evaluation, deployment — and this page — are produced by one founder directing AI agents. The method is the product: a one-person company operating a full model factory is exactly the leverage NURM packages for teams.