ChazzAI
AI roleplay and living worlds.
Beta
Models that earn their keepAI and machine learning
Training, fine-tuning, retrieval, agents — and knowing when not to.
Models trained or fine-tuned on your data, retrieval systems over your documents, agents that do real work, and the evaluation harness that proves any of it is better than what you had.
How we approach it
The work splits three ways. Training and fine-tuning, when a general model does not know your domain and no amount of prompting will teach it — full fine-tunes, LoRA and QLoRA, preference tuning with DPO or GRPO, and the dataset construction that decides whether any of it works. Retrieval, when the model does not need new weights, it needs your documents — chunking, embeddings, vector search, reranking and the citation trail that makes an answer checkable. And agentic systems, when the job is a sequence of steps rather than an answer.
The part that gets skipped everywhere else is evaluation. Before we change anything we build the harness that measures it, because "it seems better" is not a result and a model that regressed on the cases you care about will look fine in a demo.
We will also tell you when a model is the wrong tool. A rule, a query or an integration is cheaper, faster and does not hallucinate, and that answer has cost us work before.
What you get
- A measurable definition of what better means, first
- Dataset construction, cleaning and labelling strategy
- Training or fine-tuning, with the runs and results shown
- Retrieval pipeline: chunking, embeddings, reranking, citations
- Evaluation harness you keep and can re-run
- Deployment: hosted, self-hosted, or on-device
- Cost per request, measured rather than estimated
Typical timeline: Four to sixteen weeks.
What we have built with this
HumanOR
Humanize AI text and flag writing that reads machine-made.
Live
CognitiveAI
A pipeline that runs before, during and after every AI call.
Active
WebSwarm
AI testing with persona agents.
Live
Adtomaton
Brief in. Platform-ready ad out.
In Development
MXCyberSLaM
An infinite, self-playing cyber range.
Active
LMLang
A programming language designed for AI authors.
Released
AI Automation Partnership
AI and automation dropped into local businesses, as a partnership.
Active
8 of our own systems use this discipline. Every one is ours, and the ones marked open source are yours to read. The whole product line →
Client work and public source
Where this comes up
Train or fine-tune a model on your own data
A general-purpose model does not know your domain, your terminology or your documents, and no amount of prompt engineering is fixing it.
Put AI into a product without it being a gimmick
Everyone wants an assistant in the product, and most of them are a chat box bolted to a landing page that answers questions wrong with total confidence.
Build a computer vision system
You need to detect, count, read, track or identify something on camera, and the off-the-shelf products either do not fit or want your footage in their cloud.
How we work
- We talk it through
- A conversation about the problem, not a sales pitch. If we are not the right people for the job we will say so and point you somewhere better. Free, and there is no obligation after it.
- You get a written quote
- A fixed price for a fixed scope, in plain English, with a delivery date. No hourly billing that quietly doubles. If the scope changes later, you approve the change and its cost before any work starts.
- We build it in stages
- You see working software every week or two, not a status report. That means you can course-correct early, while it is still cheap to change your mind.
- You take ownership
- At handover you receive the source code, the accounts, the domains and the documentation, all in your business name. Nothing stays locked to us. You could hire anyone else tomorrow and they could pick it up.
- We look after it
- An ongoing support agreement, or on-call as you need it. Your choice, and you can change or end it whenever you like.
Last reviewed · Site changelog