Experience | Francisco Mínguez

01

Marketing, analysis, and Banco Sabadell

For about five years the work was marketing and sales analysis, and then responsibility for a business. The questions moved ahead of the tools: first a sheet that had to be rebuilt by hand, then an analysis that had to be repeatable.

Marketing and sales analysis

ProblemCampaigns and sales needed a reading someone could decide with, and the sheet had to be rebuilt by hand each time the question changed.

ContributionAt Hohe Regel I was a marketing and sales analyst: I followed campaigns and sales and turned them into that reading. At IE Law School I was a Program Account Executive, commercial work with educational programmes and with the people who might take them up. That is where I learned and applied Excel, Power Query, and Power BI.

ResultThe commercial reading could be prepared again without starting from a blank sheet.

Business reporting at Banco Sabadell

ProblemThe role asked for reports, commercial segmentation, and client relationships. An analysis that could not be repeated was not enough for that responsibility.

ContributionI was a business director and an insurance specialist. I prepared the reports, worked on the commercial segmentation, and kept the client relationships. I started using Python and R on my own, so an analysis could be run again. The specialised training came later.

ResultThe reports served the business. At the bank, that use was still a business report.

02

Studies and DeepForward

The reports showed the limit of learning a tool while the question was already due. I completed an MBA at ENEB and the master's in Data Science and Business Analytics at IMF, UCAV, and Indra. The official grade for the master's is 9.3. The master's overlapped with the AXA internship.

The final project, DeepForward, was graded 10/10. It took a question from marketing and banking and compared a model I built with a proper baseline.

The marketing degree, taught in English, is from URJC. English C1. TOEIC 910/990.

DeepForward

ProblemAfter a phone campaign, will a customer subscribe to a term deposit? The material is the public Bank Marketing dataset.

ContributionI prepared the data, studied the variables and the dimensionality, and set classical baselines before making the model more elaborate. The piece I built was a deep polynomial neural network in PyTorch. Building it meant studying how that library runs a layer, a forward pass, and a training step, and implementing the network myself.

ResultIn the documented evaluation the network stayed close to XGBoost and caught slightly more positive cases. XGBoost kept a slightly higher accuracy and AUC. What I kept is the order of work: a baseline, evaluation criteria fixed first, and complexity only when it earns its place.

Explore the case

03

AXA Seguros and EIOPA

At AXA Seguros I was on a curricular internship. The processes lived in SAS, and the size of the work was what decided the tool.

At EIOPA, in Frankfurt, the material was ESG reports. The review had to become something the team could judge.

SAS migration

ProblemThe processes I touched lived in SAS. Moving them was part of the internship, and the destination had to follow the volume.

ContributionI took part in moving them to Python and PySpark on Databricks. I also built indicators with Python, R, and Power BI, so the business of the internship could be followed with a reading that could be repeated.

ResultPython when the process fits one machine. PySpark when the volume, or the way the work is split, calls for a cluster.

ESG documents and greenwashing

ProblemThe difficulty was the form of the material: long reports, scanned pages, and language that does not leave the finding in a single field.

ContributionI built the document processing for that review. OCR for the scans, language processing (NLP), extraction, parsing, and splitting the text into passages. I compared LightGBM and PyTorch with a baseline, and I looked at anomalies with Isolation Forest and autoencoders. Alongside that, I moved SAS work to Python on Azure.

ResultThe review produced a measured recall. The team could judge the detection, and the text of the reports could be extracted and split for the reading.

04

Experian

Senior Data Scientist

At Experian I worked as a Senior Data Scientist. On the workplace-claims segmentation I was also Product Owner: technical direction, the development, and presenting the results to the clients.

The work was three pieces: processes that had to be runnable again, a segmentation that could be explained, and a spatial fit whose mathematical function already had an author.

Modernising processes

ProblemSeveral analytical processes still ran in T-SQL and SSIS. Changing one meant reopening a long script, and running it took longer than the decision that was waiting for it.

ContributionI moved them to modular pieces in Python and PySpark on Databricks: reusable classes, simple orchestration with Databricks Jobs, quality checks on the path, and continuous integration.

ResultThe migration took about four months. Execution times came down in a substantial way, and a change could be rerun and reviewed as a job.

Workplace-claims segmentation

ProblemA claims rate from a small group of firms jumps around. In the Spanish classification of activities, a CNAE class with few observations can look extreme mostly because the sample is small. One average hides the groups that behave differently.

ContributionI was the principal owner of the project. I prepared and cleaned the data and built claims indicators. Bayesian smoothing by CNAE pulls a thin class toward the wider pattern, so the rate stays usable when the sample is small. I selected and regularised variables, checked multicollinearity so a repeated signal could leave the set, and reduced dimensions with PCA. I then tried K-means configurations, grouping by centres.

ResultThe configuration I kept has nine groups of firms, each with a distinct claims pattern. The groups could be read and shown: different behaviours, rather than one average that described none of them well.

Geospatial work

ProblemThe line needed a spatial reading, and then a fit. A colleague had defined the mathematical function for the optimisation. What was missing was a way to fit its parameters.

ContributionI worked with GeoPandas and other spatial tools, and I implemented a neural network in PyTorch to fit those parameters. The function is my colleague's. The network and the fitting are the work I did. Training that network continued what I had practised on DeepForward.

ResultThe parameters of that function could be fitted, with the authorship of the function and of the network kept distinct.

05

Multiverse Computing

Fixed-term contract

The contract was for a defined period. The work covered language models, multimodal models, and computer vision. On the projects I carried, engineering and product sat together: I set priorities and kept the conversation with stakeholders.

Integration and inference

ProblemA model that runs on one machine, in one person's environment, is not yet a model the team can run again.

ContributionI integrated, packaged, and ran models across different inference backends, in containers, in the cloud, and on GPU machines.

ResultThe same model could be packaged and run again, on more than one backend, without depending on a single workstation.

Compression

ProblemThe models were heavier than the deployment could comfortably carry, and compression changes what a model does.

ContributionI compressed them by pruning, by quantization, and with tensor networks. I used knowledge distillation so a smaller model could learn from a larger one, and I worked on recovering quality after the compression.

ResultThe compressed models could be served where the original weight did not fit, and brought back toward the behaviour the use actually needed.

RAG evaluation

ProblemRetrieval plus a generated answer can look fluent when the retrieved context is thin, or when the system should have declined.

ContributionI evaluated retrieval and the answer with RAGAS, and with complementary measures: when the system should abstain, and whether the context is enough.

ResultThe evaluation showed the support for the answer, and the cases where answering was the wrong move.

06

Independent research

The question I am working on now is when a system that retrieves a passage and answers with a model should be trusted. A fluent answer can sit on a weak passage, on a citation that a newer document has replaced, or on context that is not enough.

What I want to bring together is a set of checks: retrieval that knows when it should answer, evidence that is still the right document, context that is sufficient, abstention when it is not, a comparison that is allowed to recommend nobody, and a regression check when something changes.

There is also a separate demo, on another corpus, which does not reproduce the retrieval study. It is not public at the moment.

The direction is to have these checks available together, so a change can be judged before it ships. Today the published work is the studies, the lab, and the note in the cases below.

When retrieval should answer

ProblemA retrieved passage is not automatically enough support for an answer. The decision had to be fixed before anyone saw the score.

ContributionI froze three criteria and scored them once, on a synthetic support corpus. The study decides whether the retrieved document supports an answer. It leaves the writing of the answer aside.

ResultIt met the three conditions that had been set: 17 of 20 answered questions were right, 17 of 22 answerable questions were covered, and 3 of 14 questions that should have been declined were answered. On the same 36 questions the decision was right 28 times.