CabrIA
A software-testing expert AI, trained on your tests, on your premises.
Large AI models know software testing in general, but not your applications: your conventions, your screens, your recurring failures, your history. And for many teams, sending screenshots, logs and test data to a cloud API is simply not an option. As a result, generic AI stays at the surface of your QA.
The essentials
- What: an open AI model specialized in Quality Engineering (UI, API, end-to-end, performance, mobile and UAT testing) that knows your ground truth.
- How: two memories: local fine-tuning builds the know-how into the model; the knowledge base (RAG) serves facts and lessons, with citations.
- The raw material: your real test executions (Robot Framework, Playwright, AI agents through MCP servers), captured passively and anonymized.
- Sovereignty: training and answering both run on a local AI workstation; your data never leaves your infrastructure.
- Status: an AI Cabra R&D program under construction; the evaluation bench and the capture layer are already running, demos on request, and no figure is announced unless it has been measured.
How it works
Every test execution produces experience: what worked, what broke, how it was repaired. A flight recorder captures that experience at the source, including the actions of AI agents flowing through MCP servers. A refinery sorts it, anonymizes it and only keeps what can be proven: every code example is executed before entering the corpus. The know-how goes into training; the lessons go into the knowledge base.
Two complementary memories
- Fine-tuning, the know-how. The model learns the craft: designing a test plan, writing a clean suite that follows your conventions, diagnosing a failure, repairing a locator.
- RAG, the knowledge. The knowledge base keeps what changes: known pitfalls of your applications, procedures, versions. Every answer cites its sources, and the base is updated without retraining the model.
- Either one alone hits a ceiling. The model alone forgets fresh details; a knowledge base alone cannot reason like a QA engineer. Together they make an expert that knows your ground.
On demand, everywhere you test
- In chat: ask your test-engineering questions in English or French, with answers grounded in your context.
- In the agents: before acting, an agent consults the base (known pitfalls of the screen, applicable procedure), then acts with the specialized model.
- In the IDE and CI: exposed through MCP, the open standard our tools already use; failure reports are annotated with the relevant lessons.
Where the program stands
CabrIA is an R&D program under construction, run the AI Cabra way: the evaluation bench first, proof by execution next, communication last. Progress:
- Evaluation bench: in place. Around sixty bilingual cases across five axes; several open models already measured, strengths and weaknesses quantified before any training.
- Capture: live. The flight recorder runs on several real test projects without changing a single one of their files, and secret masking is checked automatically.
- Refinery: first stages. Real incidents become training examples, scored by a model other than the one that wrote them; below the quality bar, the example is dropped.
- Knowledge base: foundation live. Local index of test documentation and keywords, search verified against real team questions.
- Next: distilled lessons, delivery to agents over MCP, then the first training run and comparative evaluation.
We will publish measurements, not promises. To discuss it for your QA team, drop us a line.
Frequently asked questions about CabrIA are gathered on the site's FAQ page.
Let's talk about your test suite
Tell us where you stand. You will get an opinion, not a brochure.