Senior AI Engineer · León, Spain · Remote

Adrian Matias

Production ML · adtech · GenAI evaluation

LLMs / RAG / agents PyTorch / transformers FastAPI · Azure · Databricks 11 years

About

Data scientist and ML engineer with 11 years of experience taking machine learning from problem framing to production. In adtech: install-prediction models behind mobile user-acquisition bidding (Appodeal), ad ranking and Bayesian bandits on the DSP side (Chartboost), persona modelling and attribution for programmatic ads (Digilant). Also fraud detection, NLP (RavenPack), recommender systems and large-scale experimentation on Spark and Databricks.

Most recent GenAI work: VET, an LLM-backed platform that evaluates enterprise software vendors, turning unstructured documents and expert-scored workshops into one assessment with cited provenance. Owned the typed output contract and the correctness gates, and built the workshop-scoring lane end to end.

Working principle: remove what is not essential.

Experience

detail
Senior AI Engineer detail Big Four firm (via Cavendish) · León, Remote · May 2026 – Present

Engagement with a Big Four professional-services firm, May – Sep 2026. One of three senior AI engineers on VET, a vendor-evaluation platform for enterprise software, with a PM and an associate product owner. Two lanes, both ending in a rich assessment: expert workshop scores over hierarchical dimensions, reconciled across evaluators, and the ingestion of unstructured documents.

Key achievements
  • Built the workshop-scoring lane end to end: parser for the domain experts' scoring templates, vendor matching, band placement, tier-weighted averaging and roll-up, reconciling workshops from different evaluators into one assessment.
  • Made provenance a property of the output: statements and metrics in the assessment quote and point back to the raw sources, so a score can be audited instead of trusted.
  • Introduced and owned schema-driven, domain-driven and spec-driven development: business requirements kept as a contract artifact that agentic development builds against.
  • Owned the typed output contract and the correctness gate set (type baselines per file and rule, a complexity ratchet, mutation tests, required status checks) and the validation workflows behind them.
  • Owned full epics end to end, contributed to every component of the architecture, and ran agentic coding on the production codebase as a measured practice: cost and turns per model tier, and, by my own count over ten weeks, 116 of my 147 pull requests merged (median 3.0 h) while I reviewed 52 % of the team's others.
Data Scientist → DS Technical Lead detail Appodeal · León, Remote · Dec 2024 – Jan 2026

Developed install prediction models for mobile user acquisition: cold start, transformers, user embeddings, transfer learning, MLOps, production pipelines, parity, experimentation and monitoring. Promoted to DS technical lead in December 2025.

Key achievements
  • Unified separate per-advertiser and per-platform pipelines and models into one custom transformer architecture. The new CPI model absorbed the click model and the advertiser-specific models. Rolled out to all Android CPI, then iOS.
  • Implemented a Bayesian evaluation framework for models trained on scarce rewards, to compare them on business metrics with estimated risk instead of a point estimate. Proposed and owned the progressive promotion funnel with milestones, challenger to champion, that grew a model's traffic share to A/B.
  • Designed and deployed cold start for advertiser onboarding in two ways: transfer learning through an affinity tower trained on a wider set of interactions, and cold-advertiser masking validated on held-out cold entities. New advertisers bid from day one, with eCPI better than curated whitelists on inventory and demand segments. Custom PyTorch.
  • Resolved parity mismatches and calibration gaps: traced discrepancies in internal transformations until the models received identical inputs offline and online, then made parity a workflow with unit tests in Databricks.
  • Proposed and designed the user embedding system, built from users' interactions with apps; delivered with transfer learning: a separately trained tower whose embeddings enter the install model as extra attention tokens.
Senior Data Scientist detail Quant AI Lab · León, Remote · Jul 2024 – Dec 2024

End-to-end ML for international clients: scope review, stakeholder alignment, model productivization.

Key contributions
  • Led the turnaround of an underperforming client ML system for utility fraud detection and ranking: found the architectural faults, fixed them, and restored accuracy, efficiency and client satisfaction.
  • Found critical faults in the data-ingestion and feature-engineering pipelines.
  • Led development of a real-time credit card fraud classifier. The open version combines gradient boosting (CatBoost) with an unsupervised anomaly score (Isolation Forest): ml-sandbox/tx_class
Senior Data Scientist detail Chartboost · León, Remote · Jun 2023 – Jun 2024

Ad ranking and reinforcement learning for in-app game advertising; improved DSP performance through large-scale experiments.

Key achievements
  • Evolved and monitored a Bayesian-bandit RL system with policy-evaluation tools; designed online/offline evaluation frameworks for it.
  • Built a two-tower neural ranking model for ads.
  • Ran post-mortems on adversarial DSP dynamics and model feedback loops.
Machine Learning Engineer detail RavenPack · León, Remote · Mar 2022 – Jan 2023

Text analytics, NLU, and semantic search on massive financial news data; MLOps and reproducibility.

Key achievements
  • Few-shot search personalization: simplified architecture.
  • Active learning PoC for learn-to-rank: uncertainty sampling on a LightGBM ranker, measured as ranking gain per acquired label against a random baseline.
  • Long-text fingerprinting for duplicate content detection.
Earlier roles & freelance detail 2015 – 2025 · ML, fraud, teaching, consulting
Numbrs · Machine Learning EngineerAug 2020 – Oct 2021Deep Learning / Active Learning for transactions; GDPR-compliant multi-country pipelines.
IEBS School · TeacherOct 2019 – Oct 2021Webinars and masterclasses on Big Data, AI, ML.
Stratio BD · Data ScientistAug 2019 – Aug 2020Propensity/churn on Spark; BERT+graph email segmentation.
DOCOMO DIGITAL · Data ScientistApr 2017 – Aug 2019Unsupervised fraud detection (behaviour clustering, outlier scoring); mentored juniors.
Digilant · Machine Learning EngineerJun 2015 – Apr 2017Consumer Persona ML redesign; Game Theory attribution.
Freelance · Freelance Data Scientist / ML EngineerMar 2022 – Jun 2025Semantic search, time series, portfolio optimization.

Selected Projects

opo — AI-compiled exam preparation opo.taoaieng.io

Live paid product for Spanish public-sector exams: official law and past papers compiled once into a cited corpus of 28 topic pages and 660 tagged official items, then a 40-item diagnostic allocates the student's hours by official weight × mastery deficit. Six proven CI gates block release; no claim ships without a citation.

GenAIKnowledge compilationEvaluationPython
The Mirror github.com/adrianmatias/the-mirror

Knowledge repository built on llm-wiki for a personal AI assistant: notes and synthesis that persist across sessions.

RAGLLM wikiAgentsObsidian
RAG Blog System ml-sandbox/ragblog

RAG over a personal blog archive with local open models: web crawling, vector embeddings and LLMs, evaluated with Ragas using separate generator and judge models. Local-first, on consumer GPUs.

RAGEmbeddingsLocal AIPython
Credit Card Fraud Classifier ml-sandbox/tx_class

Real-time fraud detection API: classical classification + unsupervised anomaly signals. Deploy-ready FastAPI, Docker, RayServe.

FastAPIDockerAnomalyRayServe
The Real MLE Challenge production ML refactor

Dev-to-production refactor: inference pipelines, a price-prediction API, Docker, reproducible structure.

MLOpsFastAPIDockerPyTorch
Document Classification seedtag-doc-classification

NLP classification with attention and transformer transfer learning. Batched prediction and GPU/CPU deployment paths.

NLPTransformersTransfer learning
not the tao — Podcast YouTube

Episodes on simplicity, consciousness, reality, machine learning, coding and sovereign AI. No guests, no ads.

PodcastMLPhilosophy

Skills

Core technologies

Python, FastAPI, Pydantic, PostgreSQL, PyTorch, Transformers, Azure OpenAI / Azure / AKS, AWS/S3, Databricks, PySpark, Pandas, SQL, Docker, MLflow, Git, GitHub Actions, AI-augmented tools (Ollama, llama.cpp, Hermes Agent)

Machine learning & GenAI

LLMs, RAG, agentic systems, LangChain / LangGraph-style orchestration, document extraction & provenance, scoring engines, deep learning (attention, embeddings, transfer learning), recommender systems, NLP, RL / Bayesian bandits, MLOps, few-shot / active learning, adtech ranking and install prediction for bidding (DSP, mobile UA)

Engineering practices

Clean code, typed contracts as source of truth, CI drift gates, OOP & FP, TDD / local e2e harnesses, reproducible pipelines, Agile / XP, production ownership

Consulting & soft skills

Multi-client consulting, stakeholder communication, mentoring, correctness gatekeeping, simplification

Education & courses

Master's in Geological EngineeringUniversity of Oviedo · 2010–2012
Degree in Mining & Energy EngineeringUniversity of León · 2012
Bachelor in Mining EngineeringUniversity of León · 2003–2007
Selected coursesCoursera / MITxClaude Code (2026) · MLOps Specialization · Deep Learning Specialization · Spark for Scalable ML · FP in Scala · Analytics Edge (MITx)

Contact

Open to senior and principal roles in ML engineering, data science and AI engineering. Remote from León, Spain.