Six years in banking and pharma, where someone always ends up asking why the system said what it said. I join your team by the day or by the hour, or take on the whole project.
B2B contracting with GLEMES FFT, directly or through your consultancy.
| Working mode | Remote · Madrid |
|---|---|
| Markets | Spain, EU, United Kingdom |
| Contracting | B2B, GLEMES FFT |
| Profiles | Infra, data, AI |
| Sectors | Banking, pharma, insurance |
§ 01 · What I can cover
I can come in as an AI engineer, a data engineer or a platform engineer. Often the gap is all three at once. A project rarely breaks inside a layer. It breaks at the seams between layers, and the seams belong to nobody.
| AI engineer | Models, RAG and agents, evaluation · PyTorch, XGBoost, LangGraph |
|---|---|
| Data engineer | Ingestion, quality, search · Spark, PostgreSQL, OpenSearch |
| Platform engineer | Cloud, deployment, CI/CD · AWS, Docker, Terraform |
I am not better than a specialist in each of the three. In at least one I certainly am not. The advantage lies elsewhere: decisions across all three are made by the same head, so the seams stop existing. When the model needs a field the pipeline does not carry, there is no meeting, there is a change.
§ 02 · Regulated sectors
In banking and pharma someone eventually asks why the system said what it said, and you need an answer. That reorders every decision. Evaluation and provenance are not added at the end: they shape chunking, schema and output format from day one.
What I do not do: demos that impress and cannot be audited. If there is no way to tell whether it works, it is not finished.
| Traceability | Every answer records which data supports it, so it can be explained months later. |
|---|---|
| Explainability | Compliance has to be able to justify every model decision. I learnt that in credit risk and fraud. |
| Model validation | Measurable evaluation from day one. Without it I cannot tell whether a change improved the system or broke it. |
| Financial crime | Alert scoring and tooling for analysts, under model validation and audit requirements. |
| Personal data | The permission filter goes inside the query, not after retrieval. |
§ 03 · Career
I started in banking, with models that had to reach production and be defended in front of the people who audit them. Financial crime is where the turn towards agents and RAG began. Since 2025 I have worked independently, through GLEMES FFT.
2025 · now
GLEMES FFT · Founder
AI consultancy. RAG and multi-agent systems from concept to production for pharma, genetic diagnostics, insurance and financial services.
2023 · 2025
ThetaRay · Senior Data Scientist
Financial crime. Alert scoring and language-model tooling for analysts, in a regulated setting with model validation and audit.
2023
Banco Sabadell · Data Scientist
Credit risk and fraud, with explainability because compliance had to be able to justify every decision.
2020 · 2023
Banco Santander · Data Scientist
Distributed ML pipelines over millions of accounts, from data to endpoint.
As GLEMES FFT I have worked with banks, pharma companies, consultancies and startups. Full career and architectures
§ 04 · Cases
Three systems I have taken into production, told by sector and without company names. Each card sums up the problem, the architecture and what it cost. The full diagram and the decision live on my personal site.
Case 01 · Systemic bank
| Problem | Alerts reviewed by hand |
|---|---|
| Architecture | XGBoost + neural net |
| False positives | -60% |
| Deployment | 3+ countries |
A single model was ruled out because the two families fail in different places and the ensemble catches what each one misses. What it costs: explainability. In banking you do not pay that in the metric, you pay it in the risk committee, and you go in with the answer ready.
Case 02 · Listed pharmaceutical
| Problem | Searching by file name |
|---|---|
| Architecture | RAG with paragraph level citation |
| Corpus | 50K+ documents |
| Evaluation | Automated, fixed question set |
Fine tuning was ruled out: in a domain where the source document changes every quarter, retraining is a recurring debt and retrieval is not. What it costs: without the paragraph citation the system is worthless, because nobody signs a regulatory report against an answer they cannot check.
Case 03 · Fintech
| Problem | Long processes, no retry |
|---|---|
| Architecture | 1 orchestrator, 4 agents |
| Volume | 10K+ operations/day |
| State | Persistent, resumable |
A single agent with many tools was ruled out because as the catalogue grows tool choice degrades and failures stop being reproducible. What it costs: latency and cost per operation. If the process fits in a function, do not reach for agents. Half the time it fits.
§ 05 · How to hire me
Always B2B: GLEMES FFT invoices, never me as an individual. The price is not on this page. It depends on the arrangement, on the length and on whether you contract directly or through a consultancy, and we cover it on the first call.
| By the day (end client) | I work inside your team, in your repository and through your reviews. Invoiced per day worked. |
|---|---|
| By the hour (consultancy) | You subcontract me for a project with one of your clients. Invoiced by the hour to the consultancy. |
| Project | A deliverable per phase that you can review before moving on. |
| Training | Sessions for technical teams on your own code: orchestration, quality, model evaluation and what to look at when something fails in production. |
On a project, the architecture is written down before anything is built, with what was ruled out and why. Redoing a decision in week three costs a conversation; in week ten it costs the project.
Forty eight checks across five blocks, with a scoring scheme. You fill it in yourself in an afternoon and come out with the state of your house in writing. It is not a brochure with the logo swapped: it is the document I use in the diagnosis phase, in full.
| Level | 1 to 4, with the scheme shown |
|---|---|
| Order of attack | What to fix first and why |
| Effort per level | In weeks of work, not in price |
| Languages | English and Spanish |
Preview: example result
Level 2. Start with Governance, not with the model.
§ 06 · Questions
Through GLEMES FFT, S.L., my company, based in Madrid. Never me as an individual. If your company needs a supplier set up, we do it with the company details and whatever paperwork your procurement team asks for.
Yes. GLEMES FFT invoices as a foreign supplier, business to business. The contract is with the company, not with me. How each engagement sits within your rules for contractors depends on the engagement, and we assess it case by case with you before signing.
Yes, yours, before the first technical conversation if needed. What I see at one client stays with that client.
You do, from the first commit. The repository lives in your organisation. There is no piece of mine running in the middle and no key I keep. If tomorrow you want someone else to carry on, they do not need my permission or my access.
Remote or on site in Madrid, depending on the project. If the engagement involves travel, we agree it in the contract.
I do not publish a minimum length: it depends on the arrangement. Length and notice are agreed in the contract and written down before we start. When I leave, the work is in a state your team can carry on without me.
GLEMES FFT invoices. By the day worked, by the hour to the consultancy, or by phase on a project. Currency and invoicing terms are agreed in the contract. We discuss rates on the call.
I only process personal data on your instructions, under a data processing agreement (Art. 28 GDPR) signed before we start, not once something is already built.
§ 07 · Contact
I reply within one working day. If you are writing from a consultancy, tell me who the end client is, even if it is only the sector.