AI implementation for businesses
I help you decide where AI adds value in your operations, integrate it with the systems and data you already use, and keep it under human control. We start from the process, not from the technology.
Updated:
There are too many tools and promises. You need a short list of use cases assessed by impact, effort and risk.
A prototype in a chat window is not a process. Integration, quality control, permissions and someone to maintain it are missing.
You worry about which data leaves for an external provider and what happens when the model gets it wrong.
Sorting emails, extracting data from documents or drafting replies is done by hand, with mistakes and delays.
Most companies do not need “an AI”; they need a specific task — reading a document, classifying a request, drafting a first version — to stop depending on manual hours. My job is to turn that task into a process with clear inputs, outputs and quality criteria, and only then decide whether a language model, a classic classification model or a simple rule is the right tool.
Technically, the usual pattern is an API integration: your system (email, form, ERP, document storage) sends the relevant information to a model, the result is validated with rules and, where appropriate, reviewed by a person before any action is taken. I work with the OpenAI and Anthropic APIs, and with alternatives when the case requires data not to leave your infrastructure.
To assess a case I need real examples (anonymised if necessary) of the input and the expected result, and to know who validates the work today. I do not need to migrate all your data or build a platform. If data is scattered or of poor quality, the initial assessment will say so and the first step will be to tidy it — that is data engineering work, not AI.
Before sending information to an external provider we define which data may leave, which is masked and which must never leave. I document the data flow so you can review it with your legal or data-protection adviser. AI proposes; a person always decides when an error has consequences for a customer, a payment or a legal obligation.
I have built Python pipelines that incorporate language models in a fintech environment (described here at a high level for confidentiality) and, as a personal project, an assistant with long-term memory and hybrid retrieval. These are not client cases, and I label them as what they are. You can see the details in the related projects.
A short document with prioritised use cases, required data, risks and a reasoned estimate of effort and benefit.
The prioritised use case running and connected to your tools (email, ERP, CRM, spreadsheets, document storage) through APIs.
Test cases built from your data, confidence thresholds and a human review flow where the risk demands it.
Architecture, decisions, API usage costs, an operations guide and training for whoever will maintain it.
I observe how the work is done today, where time is lost and what counts as a correct result. Without that, improvement cannot be measured.
I check what data exists, what state it is in and which legal or contractual restrictions apply. I rule out what is not worth doing.
I build the smallest version that can be demonstrated with your examples and evaluate it against an agreed test set.
First in parallel with the manual process, then with human review, and only if results justify it, with more autonomy.
Logging of usage, errors and costs; periodic quality review and adjustment when data or the model change.
Answer questions about manuals, policies or procedures citing the source document. Covered in depth in the RAG assistants service.
Tag invoices, contracts, tickets or emails by type and urgency and route them to the right person or system.
Turn PDFs, emails or forms into structured fields (amounts, dates, references) with later validation.
Draft replies and conversation summaries that a person reviews before sending.
Drafting reports, meeting summaries or quotes from templates and existing data.
Reconcile lists, normalise names or addresses and detect duplicates with rules plus language models.
It depends on the case. The initial assessment is usually a small, fixed-scope piece of work; integration varies with the systems involved and the level of control required. After a first conversation I give you a written estimate. Model API usage costs are added and estimated using real examples.
Only if we decide so. For each case I define what information may be sent, what is anonymised and what must never leave. Options with enterprise terms exist and, when needed, alternatives running on your own infrastructure.
It sometimes will. That is why I measure quality with a test set, apply automatic validations and keep human review for decisions with consequences. The goal is to know when and how often it fails, not to pretend it does not.
To use pre-trained language models, usually not. I need representative examples of the work and of what counts as correct. For custom predictive models, data volume and quality do matter, and we assess that first.
Yes. I work remotely with companies across Spain and other countries, in Spanish or English.
Artificial intelligencePersonal project
A language-model assistant with persistent memory and hybrid retrieval — semantic and keyword — over a graph-based store.
Data engineeringPersonal project
A personal lab that recreates an analytics engineering environment on synthetic financial-crime data, with base, core and marts layers, quality tests and orchestration.
Turn manual, repetitive tasks into scheduled, traceable and maintainable workflows.
View serviceAn assistant that answers from your documents, cites its sources and respects permissions.
View serviceReliable pipelines, clear models and quality tests so you can decide with data you trust.
View serviceForecasts and classifications honestly evaluated against simple alternatives.
View serviceA first conversation is enough to know whether it makes sense and where to start.