Artificial Intelligence Lab
AI that works
AI that works
in production
The Caixa Mágica Artificial Intelligence Lab builds production AI systems for European businesses. LLM deployments, RAG pipelines, document processing, anomaly detection and AI QA for regulated industries. Not experiments. Working software.
What is the Caixa Mágica Artificial Intelligence Lab?
The Caixa Mágica Artificial Intelligence Lab builds production AI systems for European businesses. It deploys Large Language Models on client infrastructure, builds RAG pipelines connected to internal knowledge bases, automates document processing and develops AI QA tools for regulated industries. On-premise deployment is available for all systems, with no data egress under any condition.
Capabilities
What the Artificial Intelligence Lab builds
Most AI projects fail because the engineering isn't solid enough to survive production. Our approach: architecture first, monitoring built in and a clear definition of success before writing a single line of code.
Language models and knowledge retrieval
We deploy Large Language Models on client infrastructure, in private cloud or as managed services. For organisations with sensitive data, on-premise deployment means no data leaves the environment. We also build Retrieval-Augmented Generation pipelines that connect models to internal knowledge bases and return accurate, cited responses instead of hallucinations. The difference between a generic LLM and one grounded in your own data is the difference between a contractor who is knowledgeable and one who knows your organisation.
Automation, detection and learning
Document processing, anomaly detection and process automation address the operational work that consumes engineering capacity. In each case, the Artificial Intelligence Lab builds systems that generate actionable outputs with manageable error rates. Continuous learning pipelines are standard practice here, because AI systems that are not regularly retrained degrade quietly and stop delivering value.
LLM Deployments & RAG
We deploy open-source models such as Llama and Mistral on client infrastructure, on private cloud or as managed services. On-premise deployment means no data leaves your environment under any condition.
RAG pipelines connect models to your internal knowledge, policy documents, technical manuals, contracts, regulatory filings and return accurate, cited responses. A model grounded in your own data is a fundamentally different tool from a generic assistant.
Document Processing
We automate the extraction, classification, and routing of information from unstructured documents: contracts, invoices, regulatory filings, technical reports. Classical NLP combines with modern LLM pipelines for documents that include images, tables or non-standard formats.
The output feeds into your existing workflows as structured data. No separate system to manage, no manual review queue to maintain.
Recommendation Systems
Content-based and collaborative filtering systems that adapt to user behaviour in real time. These are built for retail, media and enterprise environments where the recommendation logic must be explainable and auditable. A black box that produces results nobody can account for is a liability, not an asset.
Anomaly Detection
Statistical and machine learning models that surface unusual patterns in operational data: fraud indicators in financial transactions, performance degradation in infrastructure, deviations in manufacturing quality.
An anomaly detection system that fires constantly is worse than no system at all. The engineering challenge is calibration as much as detection and we treat it that way from the start.
Process Automation
Automation that goes beyond rules and triggers. We build systems that read unstructured inputs, make classification decisions, route work to the right team, and escalate exceptions to humans. Operational cost drops without removing human judgment from decisions that genuinely require it.
Continuous Learning
AI systems degrade when the world changes and the model doesn't. We build feedback loops that capture user behaviour and operational outcomes, then use them to retrain models over time. This is standard practice in the Artificial Intelligence Lab. The main reason most AI projects stop delivering value after six months is that nobody built the retraining loop.
Proprietary product
Qualigentic, AI QA for Regulated Industries
What it does
Our AI QA platform reads requirements from Jira, ALM or Confluence, generates production-ready test scripts for Selenium, Cypress, Playwright and Robot Framework and maintains the test suite autonomously. In particular, it classifies failures, refactors tests and tracks coverage without manual intervention.
Compliance and data residency
For financial services clients, it produces DORA, Solvency II and PSD2-compliant audit evidence chains with signed logs, RBAC controls, and configurable retention policies. On-premise deployment is available from the entry tier. As a result, fine-tuning weights never leave client infrastructure under any condition.
Why it is not a generic AI assistant
Generic AI assistants suggest test code. Qualigentic runs tests, monitors results, maintains suites and produces the compliance documentation a regulatory audit requires. That gap is substantial and grows wider in regulated industries where audit evidence is not optional.
Learn about QualigenticRequirements to tests, automatically
Reads from Jira, ALM, Confluence. Generates Selenium, Cypress, Playwright, Robot Framework scripts. No proprietary runtime.
On-premise, zero data egress
Self-hosted open-source models (Llama, Mistral). Fine-tuning weights never leave client infrastructure. Banking, healthcare, government.
Audit-grade compliance evidence
DORA, Solvency II, PSD2. Full traceability from requirement to test to execution. Signed audit logs, RBAC, configurable retention.
Autonomous suite maintenance
Classifies failures, refactors broken tests, identifies coverage gaps without manual intervention.
Approach
How the Artificial Intelligence Lab works
Every AI engagement starts with a data audit and a feasibility assessment. If your data isn't ready to support the system you want to build, we say so before we start.
Three principles behind every AI project
Data readiness is audited in week one, because most AI projects fail before the engineering begins. In addition, all models, training data, and pipelines are owned by the client from day one. For regulated industries, auditability is an architectural requirement, not a feature retrofitted after the regulator asks for it. These three principles shape every engagement the Artificial Intelligence Lab takes on.
Data-first
Data readiness is audited in week one. Most AI projects fail because the underlying data was never adequate. Surfacing this before engineering begins is far cheaper than discovering it six months in. If the data isn't ready, you get an honest assessment of what it would take to fix it.
Open by default
You own your models, training data and pipelines. Open-source frameworks are the default, so there is no dependency on us or any proprietary platform to keep the system running after the engagement ends.
Auditable by design
For regulated industries, AI systems need traceable decisions, signed logs and documentation structured to satisfy a compliance review. Auditability is built in as an architectural requirement from day one, not added after the regulator asks for it.
Technology
Technology stack
The frameworks and infrastructure our AI engineers use in production, chosen for reliability, openness and suitability for regulated environments.
Open-source first, cloud-agnostic
The Artificial Intelligence Lab works primarily with open-source frameworks, which means clients own their models and are not tied to any single cloud provider. In practice, this covers the full ML lifecycle: model selection, training, inference, orchestration and data storage. For regulated clients who need on-premise deployment, every component in this stack runs within their own infrastructure.
LLMs & AI Frameworks
ML & Data Science
Infrastructure & Orchestration
Vector Stores & Databases
FAQ
Artificial Intelligence Lab — frequently asked questions
Questions from clients exploring AI for the first time and from those who have been through a project that never shipped.
What is the difference between using ChatGPT and building a custom AI system?
Generic AI assistants are useful for individual productivity tasks. A custom AI system integrates with your data, your workflows and your infrastructure and produces results that are consistent, auditable and specific to your organisation. For enterprise use, the gap is similar to the one between using a public search engine and building an internal knowledge system connected to your own data.
Can AI models be deployed on our own infrastructure?
Yes. For clients with data residency requirements, which is standard in banking, healthcare, insurance and government, we deploy models on-premise or in private cloud environments. Open-source models such as Llama and Mistral run entirely within your infrastructure, with no data leaving under any condition.
How do you ensure AI outputs are accurate and reliable?
Through validation frameworks, human-in-the-loop review at appropriate points and continuous monitoring of model performance in production. We build systems with explicit failure modes. The system knows when it does not know, and escalates to a human rather than producing a confident wrong answer.
What data do you need to get started?
It depends on the problem. Document processing needs a representative sample, typically 200 to 500 examples, to assess variability and structure. Recommendation systems need historical interaction data. Anomaly detection needs baseline operational data with some labelled anomaly examples. We assess data readiness in the first phase of every engagement and are direct about what is and isn't feasible.
How long does an AI project take?
A focused automation project, such as document classification, a RAG-based knowledge system or a recommendation API, typically takes 6 to 12 weeks from discovery to production. More complex systems with custom model training take 3 to 6 months. Qualigentic, our AI QA platform, can be deployed in 2 to 4 weeks depending on the environment.
Do you work with companies that already have a data science team?
Yes. We work alongside existing data science and ML teams regularly, providing production engineering capacity, MLOps infrastructure, or compliance expertise for regulated industries. The Artificial Intelligence Lab is most useful as an extension of your team, not a replacement for it.
Have an AI problem to solve?
Tell us what you are trying to build or automate. We respond within one business day.
Let's talk