AI Engineering · Models you own
We train, release, and run specialised AI models you own.
Naderu is an AI-models company. Not a chatbot, not an API wrapper — the engineering that turns a foundation model into a dependable, domain-specialised model your organization owns and can deploy anywhere. Engineered, not merely prompted.
What we do
Three things, done properly.
We are deliberately narrow. The product is the model — and the engineering that makes it trustworthy.
Train
Fine-tune, adapt (LoRA), distill, and quantize foundation models into specialised ones. Data engineering and evaluation are first-class, not afterthoughts.
Release
Publish models with real model cards, benchmarks, and clear licensing. Every release names its upstream foundation model — no black boxes.
Serve
Consulting, customization, deployment, and operations — the engineering around models, sold as services. You own the result.
Engineered, not prompted
A prompt is one layer. A model is a lifecycle.
Real enterprise AI is data, training, evaluation, deployment, and monitoring — versioned and reproducible end to end.
Evidence over claims
Every model ships with its receipts.
No model reaches release until it passes its evaluation gate. Cards and benchmarks are part of the product.
naderu-geek-py-0.5b
Our first public portfolio model: a compact LoRA fine-tune of
Qwen2.5-Coder-0.5B-Instruct
(Apache-2.0) that imprints a clean, commented-Python style and a recognisable
naderu-geek identity. It demonstrates our pipeline — data → training →
evaluation → release — end to end. It is a demonstration piece, not a capability
claim: coding ability tracks the base. It passed a qualitative smoke eval
(identity recognised · coherent Python), with the card and provenance stated plainly.
naderu-loom-7b
An offline interactive-fiction narrator: given a game state and the player's action, it returns a single JSON turn (scene + 2–4 choices + state changes) an app can render and apply — a game master that runs on-device. A QLoRA fine-tune of Mistral-7B-Instruct-v0.3 (Apache-2.0), trained on Apple Silicon. Unlike a smoke test, it ships a quantitative eval gate — and passes it: 100% valid JSON and 0 world-state violations across 34 scripted turns.
hunter-crypto-7b
The first model of the Hunter family — a local cryptography-attack specialist for authorized security and CTF work. Given a weak or misconfigured construction, it names the flaw and emits a runnable attack script that recovers the secret, fully offline on Apple Silicon. A QLoRA fine-tune of Qwen2.5-Coder-7B-Instruct (Apache-2.0). Its gate is execution-graded — the attack recovers the secret or it does not — and it passes: 100% solve rate on the in-distribution held-out set and 97% on an independent-method set. A narrow specialist, authorized use only.
- model
- naderu-geek-py-0.5b
- foundation
- Qwen2.5-Coder-0.5B-Instruct (Apache-2.0)
- task
- Python coding assistant · identity + style
- method
- LoRA (r=16, α=32) · merged · fp32→bf16
- data
- 39 examples · license-clean · in-repo
- license
- Apache-2.0 · inherits upstream terms
- status
- passed eval gate
| check | outcome |
|---|---|
| identity recognised | yes |
| coherent, runnable Python | 3/3 |
| held-out identity denial | pass |
| held-out style transfer | pass |
| provenance & licence stated | yes |
Honest scope — this is a qualitative smoke eval, not a capability claim: naderu-geek's coding ability tracks its base. Capability-claiming releases carry reproducible quantitative benchmarks tied to a versioned eval suite.
- model
- hunter-crypto-7b
- foundation
- Qwen2.5-Coder-7B-Instruct (Apache-2.0)
- task
- Crypto-attack scripts · authorized security / CTF
- method
- QLoRA (r=8) · MLX · merged bf16
- data
- synthetic · gate-validated · 6 categories · license-clean
- license
- Apache-2.0 · inherits upstream terms
- status
- passed eval gate
| held-out tier | base → trained |
|---|---|
| in-distribution solve_rate (48) | 0.333 → 1.000 |
| valid-script rate (48) | 0.333 → 1.000 |
| independent-method (36) | 0.250 → 0.972 |
| OOD-hard (24 · reported) | 0.000 → 0.125 |
Capability claim, backed — unlike a smoke eval, hunter-crypto ships reproducible, execution-graded benchmarks tied to a versioned suite. It is a narrow specialist: strong on its six trained attack families, not a general cryptanalysis engine (see the OOD tier).
Own it · Deploy anywhere
Your model. Your infrastructure. Your data.
Privacy by default. Run where compliance and cost demand — we don't hold your model hostage.
Services
Engineering, delivered as a partnership.
Consulting
Where a specialised model beats a general API — with the roadmap and ROI to prove it.
Customization
Fine-tune, distill, and quantize a foundation model to your domain and data.
Deployment
Stand the model up where you want it — cloud, private, or edge — and hand you the keys.
Operations
Monitoring, re-evaluation, and continuous improvement of models in production.
Build your model
Come build an intelligent capability you own.
Tell us the problem. We'll tell you whether a specialised model is the right tool — and if so, engineer it end to end.