Causal effects, not just predictions
Estimating heterogeneous treatment effects (CATE/HTE) with doubly-robust and orthogonal methods, and using them to build decision policies rather than stop at a point prediction.
Staff Data Scientist · LATAM Airlines
Bruno Vilela Quirino
I work at the boundary between causal inference and machine learning: estimating the effect of actions, learning policies from those estimates, and building systems that decide under uncertainty. In production, that means contextual bandits and treatment-effect models inside large-scale personalization and marketing platforms. Alongside that, I'm building an independent research track, one paper at a time.
About
From engineering to decision science
I started as a software engineer, moved into data science, and have spent the last few years narrowing in on the layer where statistics turns into action: causal inference, contextual bandits, and policy learning. Most of my day-to-day is building and operating decision systems in production — pricing, personalization, marketing — but I treat that work as a source of research questions, not just delivery. Competitions, an undergraduate thesis in space-plasma physics, and problems from work have each, in turn, become a methods question worth writing up properly.
A tabular Q-learning agent training live in your browser — no server, no dataset, just the Bellman update running client-side.
Episode 0 · ε = 1.00 · best path: —
Research
Causal inference and decision-making under uncertainty
My focus is the gap between a good model and a good decision: how to move from an estimate to a policy that can be trusted, measured, and defended once it's live.
Bandits that keep their guarantees
Contextual bandits and Thompson Sampling for sequential decisions, with an eye on what breaks off-policy evaluation once a system starts learning from its own exploration.
Geometry-Aware Flight-Time Prediction from Convective Polygon Intersections
Manuscript in preparationA gradient-boosted regressor for realized takeoff-to-landing flight time that treats convective weather as geometry — segmenting storm imagery into georeferenced polygons and spatially joining them against radar-derived flight tracks, instead of collapsing weather into airport-level scalars. Grew out of a 2nd-place finish in the LATAM/ITA/DECEA air-traffic data science challenge.
Method and result
- Fuses air-traffic movement logs (BIMTRA), EUROCONTROL CAT-62 radar tracks, METAR/METAF weather reports, and image-derived convective polygons into a single LightGBM pipeline.
- Explicit spatial joins between flight tracks and georeferenced storm polygons — the geometry of convection as a first-class feature, not an image fed to a CNN.
- 286,080 flights, 73 engineered predictors; 5-fold cross-validated RMSE of 363s (473s held-out), roughly halving error on convective flights versus a no-imagery baseline.
Selected Experience
Building decision systems end-to-end
LATAM Airlines
Staff Data Scientist- Own NBO platform architecture on GCP Vertex AI; act as design authority across squads.
- Build uplift modeling pipelines powering ancillary upsell and lifecycle marketing triggers.
- Design and deploy contextual bandit services for experimentation and journey orchestration.
- Deliver product recommendation engines (collaborative filtering, content-based) integrated with personalization APIs.
- Develop customer lifetime value segmentation and monitoring dashboards linked to commercial planning.
- Lead architecture reviews, privacy controls, and MLOps standards (CI/CD, registry, monitoring) to reduce time-to-production risk.
Kumulus
Senior Data Scientist- Delivered RAG pipelines and GenAI chatbots spanning web and WhatsApp, covering document understanding, demand forecasting, and computer vision briefs.
- Applied contextual bandits and reinforcement learning to guide chatbot flows and conversion prompts against user-specific variables.
- Orchestrated Databricks/Spark ETLs with Azure/AWS integrations, leveraging LangChain retrieval workflows and production observability.
Plusoft
Data Scientist- Executed large-scale segmentation (LTV, churn tendencies), geomarketing, and time-series models.
- Implemented PySpark pipelines for data quality, modeling, dashboards, and API integrations.
Vertem
Data Scientist- Developed anti-fraud risk modeling, segmentation, and CLV initiatives with Azure/MLflow MLOps.
- Drove stakeholder storytelling and model monitoring practices.
Earlier Full-Stack Engineering (Intern/Junior)
- Built full-stack web apps using PHP (Laravel/CodeIgniter), MVC, JavaScript/jQuery, HTML/CSS/Bootstrap.
- Delivered REST APIs, Docker setups, and foundational CI/CD for modular microservices.
Skills
Research and engineering toolkit
Causal Inference & Decision Science
- Causal inference: CATE/HTE estimation, doubly-robust methods
- Contextual bandits: Thompson Sampling, off-policy evaluation
- Recommenders (NBO); uplift modeling
- A/B testing & DoE; interpretable linear/logistic models
ML & MLOps
- Reinforcement learning (contextual bandits, Q-learning)
- RAG/Embeddings; Vector DB; Feature stores; Observability
- CI/CD for ML; Model registry; Lineage; Monitoring & drift; IaC
Platforms & Engineering
- GCP Vertex AI; Azure (Data/AI); AWS
- Microservices; Event-driven; API-first; Architecture reviews; Well-Architected; Privacy-by-design
- Spark/Databricks; Python, Scala; REST APIs; earlier full-stack background (PHP/Laravel, JS)
Recognition
Awards & Certifications
Awards
Certifications
- Azure Fundamentals; Azure AI-900
- Oracle AI
- Databricks Data Scientist
- GCP Vertex AI
- AWS SageMaker
Projects
Open source & research
COSMIC - Python package for magnetic turbulence analysis
RepositoryA reusable library inspired by ESA Cluster mission data. Originated from an undergraduate thesis on magnetic reconnection and extreme-event statistics, now released as a tested package with a clear API.
Key features
- Ingestion and preprocessing for .dat and CEF files, consolidating the logic from original notebooks.
- Current density and gradients via the curlometer method.
- Turbulence metrics: PVI, magnetic volatility, Gaussian-kernel smoothing.
- Robust time-series cleaning and normalization: outlier detection, magnetic energy, and one-line resampling utilities.
- Spectral diagnostics: total/parallel/perpendicular PSD, temporal and spatial correlations, high-order structure functions.
- MVA, multi-point timing analysis, field-line curvature estimation, helicity, tetrahedron quality factors.
- Extreme value analysis tools: POT, declustering, and associated visualizations.
- Automated tests with pytest for reproducible results.
MarkowitzPortfolioOPT
RepositoryProduction-oriented implementation and backtesting utilities for portfolio optimization with Brazilian equities.
DMC20
RepositoryData Mining Cup solution pipeline; documented approach and experiments.
HashCode problems
Documented solutions with complexity notes and heuristic strategies.
Selected repositories: morepizza_hashcode, hashcodeDataCenterOptimization, adventofcode20.
Writing
Notes on methods and results
DMC 2022: SVD x Markov Chain
Shows how combining SVD-based user and category similarity with Markov chains forecast weekly repeat purchases for DataMiningCup 2022, using Bayesian tuning to hit 92% accuracy within hardware limits.
Identifying Extreme Events with the Dragon-King Test
A practical walkthrough on detecting heavy-tail anomalies; statistical assumptions, implementation notes, operational caveats, and visualization tips.
Contact
Let's talk
Happy to talk about causal inference, bandits, decision systems, or possible collaborations — research or applied. Reach out via the form or connect on LinkedIn.