Projects

Selected projects

Applied AI work from technical strategy to production delivery.

A portfolio of industry and research projects translated into senior Data Scientist, Applied Scientist, and Machine Learning Engineer language: model design, experimentation strategy, scalable data pipelines, computer vision, statistical rigor, and stakeholder-facing delivery.

Industry Projects

E-commerce recommendations

Recommendation Systems at Mercado Libre

Production ranking model serving millions of users across Argentina, Brazil, and Mexico.

Ranking graph

Primary data science owner for an online ranking model used by dozens of internal clients at LATAM scale. The work combined PyTorch model development, production feature-store integration, embedding changes, CUDA and Fury pipeline optimization, MLflow experiment tracking, ONNX artifact validation, and monitoring in Datadog and Looker. I guided controlled production experiments, communicated trade-offs and risks, and supported release decisions using statistical evidence and business KPIs.

PyTorchRankingRecommendersMLflowONNXA/B testingProduction ML
Read more
  • Challenge: improve model behavior and release safety in a high-traffic recommendation environment.
  • Contribution: tripled training-data capacity through pipeline, framework, and CUDA optimization.
  • Transferable skills: experimentation, model governance, cross-functional delivery, monitoring, and production ownership.

Applied data science leadership

Data Science Lead at IThreex Global

Eight-person team delivering client-facing AI and analytics products.

Platform map

Led an eight-person data science team across public revenue, agriculture, retail, tourism, and international trade. I defined technical roadmaps and OKRs for the Molibdeno AI platform, designed the core Python library for reusable ML workflows, mentored and onboarded data scientists, and acted as technical contact for client projects. Delivery included tax-revenue forecasting, payment-behavior analysis, customer segmentation, computer vision for cattle-weight estimation, and a LangChain RAG support API.

PythonForecastingComputer visionRAGRoadmapsOKRsTeam leadership
Read more
  • Challenge: turn ambiguous client needs into reusable, maintainable data products.
  • Contribution: provided technical direction, mentoring, client reporting, and reusable platform foundations.
  • Transferable skills: AI solution design, stakeholder management, delivery planning, and technical leadership.

Research Engineering Projects

Real-time event detection

TOROS: Gravitational-Wave Counterpart Detection

Real-time computer vision pipeline for rare-event discovery.

Alert pipeline

Developed software and machine learning workflows for identifying visible-light signals associated with gravitational-wave alerts during the LIGO/Virgo O1 and O2 follow-up campaigns. After an alert, the pipeline processed telescope images within hours, compared new observations against references, detected candidate changes, ranked them statistically, and supported follow-up decisions across observatories. TOROS responded to three O2 alerts, including GW170104 and GW170817, translating directly to anomaly detection settings where speed, false-positive control, and human-in-the-loop validation matter.

4,853 labeled examples Random Forest F1 = 0.89 39 candidates from 2,375 rejected bogus detections
PythonComputer visionRandom ForestDifference imagingReal-time alertsHuman-in-the-loop ML
Read more
  • Challenge: find rare signals in noisy images under tight follow-up windows, while controlling subtraction artifacts and false positives.
  • Contribution: difference imaging, candidate detection, statistical prioritization, and validation workflow design using 2,414 bogus artifacts and 2,439 injected transients.
  • Scale: three observing facilities in Argentina and Chile; 26 galaxies followed during the first two nights after GW170817.
  • Source: TOROS O2 paper.
  • Industry analogues: fraud detection, medical alerting, security monitoring, and manufacturing anomaly detection.

Data-intensive pipeline framework

Corral Framework

SQL-backed Python framework for reproducible data-processing pipelines.

Pipeline engine

Contributed to Corral, a Python framework for building data-intensive astronomical pipelines around an MVC architecture on top of SQL relational databases. Corral supports custom data models, staged processing, communication alerts, multiprocessing, distributed computing, and unit-test-based quality metrics. In industry terms, this is platform engineering for reproducible data workflows: structure the data model, orchestrate transformations, track quality, and make complex pipelines maintainable by teams.

SQL-backed MVC architecture Multiprocessing and distributed computing 7-author public framework paper
PythonSQLPipeline orchestrationTestingDistributed processingData models
Read more
  • Challenge: standardize complex data-processing workflows while preserving flexibility for different scientific pipelines.
  • Contribution: framework design, public code, pipeline patterns, quality metrics, and reproducible workflow practices.
  • Source: public GitHub code and a 19-page Astronomy and Computing paper.
  • Industry analogues: ML/data platform engineering, ETL orchestration, analytics infrastructure, and quality-controlled batch processing.

Public-interest analytics

Arcovid19 Decision-Support Platform

Open-data pipelines and dashboards during the COVID-19 pandemic.

Dashboard

Co-developed Arcovid19, a multidisciplinary open-data initiative integrating epidemiological datasets, statistical analyses, and interactive visualizations during the COVID-19 pandemic. The project was organized around five freely available products and work axes, including automated official-data curation, reconciliation workflows, error calculation, epidemiological scenario generation, benchmarking, and reporting. Its value was not only modeling, but also building transparent workflows that could support public-facing interpretation and institutional decision-making under uncertainty.

5 open products/work axes Brooks spreadsheet-ingestion tool Official-data curation and reconciliation
ETLOpen dataForecastingDashboardsNoisy data ingestionPublic health
Read more
  • Challenge: integrate fast-changing public data into reliable analytical products despite spreadsheet errors, duplicate records, and source inconsistencies.
  • Contribution: data engineering, statistical analysis, visualization, reproducibility, and scientific communication.
  • Engineering detail: Brooks was designed for rapid epidemiological loading from spreadsheets while tolerating errors and duplicates.
  • Industry analogues: healthcare analytics, business intelligence, executive dashboards, and risk monitoring.

Computer vision at scale

Automatic Galaxy Detection in Infrared Images

Object detection in extremely crowded and noisy image data.

Image detector

Designed computer vision methods for detecting galaxies hidden behind dense foreground stars in near-infrared survey images. In plain terms, the task was to identify faint extended objects in images where the background is crowded, noisy, and uneven. This work contributed to VVV/VVVX catalog efforts spanning hundreds to thousands of square degrees, using CNNs on image data and XGBoost on photometric and morphological features. The same methods map naturally to medical imaging, satellite imagery, and industrial defect detection.

VVV: 562 sq. deg. VVVX: 1,700 sq. deg. CNN + XGBoost classifiers
Computer visionCNNXGBoostFeature engineeringQuality controlCatalog generation
Read more
  • Challenge: distinguish faint extended objects from dense backgrounds and artifacts.
  • Contribution: object detection, morphological analysis, ML classification, and automated validation workflows.
  • Scale: the Southern Galactic disc catalogue covered 1,080 square degrees and contains 167,559 galaxy candidates.
  • Validation: 14% of catalogue objects were confirmed through visual inspection or cross-match with previous catalogues.
  • Transferable skills: noisy-image modeling, defect detection, visual QA, and scalable labeling support.

Time-series machine learning

Variable Star Classification

Supervised learning from irregular time-series observations.

Time series

Developed machine learning methods for classifying variable objects from repeated observations over time. The technical problem was similar to many industrial time-series tasks: observations are irregular, noisy, incomplete, and must be transformed into robust features before classification. The project involved signal processing, feature extraction, supervised learning, and model evaluation, with emphasis on reliability rather than a single metric. It also required deciding which features were stable enough to generalize across instruments, observing conditions, and object classes.

Time seriesSignal processingSupervised learningFeature extractionModel evaluation
Read more
  • Challenge: learn from irregular sequences with missing observations and variable signal quality.
  • Contribution: feature design, classifier development, validation, and scientific interpretation.
  • Industry analogues: predictive maintenance, IoT analytics, finance, and customer behavior modeling.

Large-scale catalog engineering

VVV Near-Infrared Galaxy Catalogue

Automated catalog construction from millions of survey detections.

Data catalog

Contributed to one of the largest catalogs of galaxies hidden behind the Galactic Plane by processing infrared detections from the VISTA Variables in the Via Lactea survey and its VVVX extension. The engineering work included extraction, filtering, automated validation, metadata organization, and statistical characterization. The project is best understood as a data-platform problem: converting noisy raw detections into a reliable, searchable, documented dataset suitable for downstream analysis.

167,559 galaxy candidates 1,080 sq. deg. Southern catalogue 1,003 new Northern galaxies
Data engineeringPythonDatabasesETLMetadataData quality
Read more
  • Challenge: transform massive observational output into a trusted structured catalog while replacing large parts of manual inspection with scalable ML and quality control.
  • Contribution: automated extraction, catalog generation, quality validation, and statistical summaries.
  • Discovery: the Northern Galactic disc catalogue identified 1,003 new galaxies, with only two previously known.
  • Industry analogues: product catalogs, data warehouses, metadata platforms, and master-data workflows.

Regression under uncertainty

DEEPz Photometric Redshift Estimation

Machine learning for distance estimation from noisy observations.

Regression

Worked on machine learning models for estimating galaxy distances from photometric measurements, a lower-cost alternative to more expensive direct measurements. Technically, this is a regression problem with noisy, incomplete, high-dimensional inputs and uncertainty that must be communicated honestly. DEEPz produced photometric redshifts for more than 1.3 million galaxies over 50.38 square degrees, using deep-learning techniques such as transfer learning, mixture density networks, autoencoders, and ensemble models.

1,341,559 galaxy predictions 50.38 sq. deg. survey area 20-50% smaller scatter for faint galaxies
Deep learningRegressionUncertaintyCalibrationTransfer learningEnsembles
Read more
  • Challenge: infer continuous quantities from indirect measurements with known noise and bias.
  • Contribution: model development, uncertainty-aware validation, calibration, and robust feature handling.
  • Performance: for faint galaxies, DEEPz achieved 20-50% smaller scatter than BCNz2 in several fields; earlier work reported a 50% reduction in σ68 scatter at i_AB = 22.5.
  • Industry analogues: price prediction, demand forecasting, credit scoring, and risk modeling.

Spatial analytics

Large-Scale Structure of the Universe

Statistical modeling of massive three-dimensional spatial datasets.

Spatial map

Developed statistical methods to study how galaxies are distributed across very large volumes of space using observational data and simulation outputs. Stripped of domain jargon, the work involved spatial pattern discovery, clustering, simulation, Bayesian inference, and model comparison on massive 3D datasets. I worked with SDSS observational data and the Millennium Simulation, whose public database represents 10^10 simulated particles in a 500 h^-1 Mpc box. The transferable value maps to geospatial analytics, network analysis, and spatial intelligence platforms.

SDSS + Millennium data 10^10 simulated particles Jackknife uncertainty estimation
Spatial statisticsBayesian inferenceMonte CarloCorrelation functionsRPython
Read more
  • Challenge: extract reliable structure from sparse, biased, and high-dimensional spatial data.
  • Contribution: statistical methodology, simulations, inference workflows, and publication-quality validation.
  • Methods: correlation functions, jackknife uncertainty estimation, friends-of-friends/percolation, and minimal spanning trees.
  • Industry analogues: location intelligence, logistics, market mapping, and network analytics.

Unsupervised discovery

Cosmic Void Detection

Finding low-density regions in massive 3D datasets.

Clustering

Designed algorithms to discover large low-density regions inside complex three-dimensional point-cloud data. In industry language, this is unsupervised structure discovery: estimating density, finding meaningful segments, handling geometry, and validating whether detected patterns are real or artifacts. The analysis compared observational and simulated datasets, including 252 voids in SDSS and 4,015 voids in the full Millennium simulation box, with radii from 6 to 24 h^-1 Mpc.

252 SDSS voids 4,015 simulation voids Density below 10% of cosmic mean
Unsupervised learningClusteringDensity estimationGeometryScientific computing
Read more
  • Challenge: identify meaningful low-density regions without labels or simple ground truth.
  • Contribution: algorithm design, validation, simulation, and interpretation of detected structures.
  • Method detail: used density-contrast thresholds below -0.8 and total density below 10% of the cosmic mean.
  • Transferable skills: segmentation, anomaly discovery, spatial modeling, and unsupervised validation.

Experimentation and inference

Galaxy Spin Alignment Analysis

Hypothesis testing over large observational datasets.

A/B testing

Performed statistical analyses of orientation patterns to test whether measured directions correlate with the surrounding environment. The business translation is experimentation under observational constraints: define a hypothesis, construct a reliable null model, estimate significance, and avoid over-interpreting noisy correlations. The work required Monte Carlo simulations, correlation analysis, uncertainty quantification, and careful communication of statistical evidence. This kind of reasoning is directly useful when product or model changes appear promising but the evidence is subtle.

Hypothesis testingMonte CarloCorrelation analysisBayesian statisticsExperimentation
Read more
  • Challenge: separate weak statistical signals from random alignments and selection effects.
  • Contribution: significance estimation, simulation-based validation, and evidence communication.
  • Industry analogues: A/B testing, causal inference, marketing analytics, and model-impact analysis.

Entity resolution

Cross-Matching Astronomical Catalogs

Matching millions of records across heterogeneous data sources.

Record linkage

Developed scalable methods for matching objects across catalogs produced by different instruments and surveys. The technical problem is entity resolution: multiple sources describe overlapping real-world objects, with different errors, formats, missing values, and uncertainty. The work involved approximate nearest-neighbor search, large database integration, reconciliation rules, and quality checks. These are directly relevant to customer identity resolution, CRM integration, master data management, and data fusion.

Entity resolutionNearest neighborsDatabase integrationRecord linkageData quality
Read more
  • Challenge: join uncertain records at scale when exact keys do not exist.
  • Contribution: matching algorithms, reconciliation logic, validation, and scalable search design.
  • Transferable skills: identity resolution, deduplication, metadata handling, and source integration.

Reproducible data platforms

Automated Scientific Catalog Construction

ETL pipelines for validated, publishable datasets.

ETL flow

Designed reproducible pipelines for extracting, validating, enriching, and publishing structured datasets from heterogeneous observational sources. The work maps closely to modern analytics engineering: define data contracts, automate transformations, validate output quality, preserve metadata, and make results reproducible for downstream users. The scientific setting increased the rigor because published datasets must remain auditable long after the original analysis is complete.

ETLPipeline automationMetadataValidationReproducibilityDatabase design
Read more
  • Challenge: turn heterogeneous raw inputs into durable, auditable data products.
  • Contribution: pipeline design, validation rules, metadata handling, and publication workflows.
  • Industry analogues: enterprise ETL, analytics platforms, data governance, and data products.

Visual analytics

Scientific Visualization

Visual tools for exploration, communication, and uncertainty.

Visual analytics

Created visual analytics tools for exploring high-dimensional scientific datasets and communicating uncertainty to technical and non-technical audiences. The work involved exploratory analysis, statistical storytelling, plots that reveal model behavior, and visual checks that make data quality problems visible. This is the same communication layer needed in product analytics, BI dashboards, model monitoring, and executive reporting. The emphasis was on helping people make better decisions from complex data, not just producing attractive figures.

EDAVisualizationData storytellingUncertaintyDashboardsCommunication
Read more
  • Challenge: make complex models and uncertainty understandable without oversimplifying them.
  • Contribution: exploratory visualization, communication assets, and visual model diagnostics.
  • Industry analogues: executive dashboards, BI visualization, product analytics, and model reporting.

Reusable software

Open-Source Scientific Software

Reusable libraries, documented pipelines, and collaborative tools.

Library

Designed and contributed to reusable scientific software supporting statistical analysis, machine learning workflows, reproducible research, and collaboration. Projects include HEARSAY, AEGIS, PINNACLE, PROPERIMAGE, GriSPy, and related repositories. Across these tools, the emphasis was repeatability, documentation, automation, package structure, and long-term usability by other researchers. This is directly comparable to internal ML platforms, data science libraries, and developer tools.

PythonOpen sourceDocumentationTestingReusable librariesML workflows
Read more
  • Challenge: make research code reusable, documented, and robust enough for collaborators.
  • Contribution: library design, automation, documentation, software packaging, and reproducibility practices.
  • Transferable skills: platform thinking, developer experience, maintainability, and collaborative engineering.

Probabilistic modeling under uncertainty

OTHER: Modeling Contact Between Civilizations

Simulation-based reasoning for rare events with extreme uncertainty.

Simulation

Developed and simulated a computational model for estimating probabilities of causal contact between civilizations in the Milky Way. The model was intentionally simple and interpretable, organized around three core questions: how rare civilizations are, how long they last, and how far their signals can be detected. The project combined astronomy, anthropology, biology, philosophy, architecture, and theology/religion, requiring probabilistic reasoning and clear communication across disciplines.

3-parameter interpretable model Interdisciplinary research team Simulation-based public communication
Monte CarloSimulationProbabilistic modelingUncertaintyCommunicationInterdisciplinary work
Read more
  • Challenge: reason quantitatively about a rare-event problem where direct empirical data are extremely limited.
  • Contribution: computational modeling, simulations, interdisciplinary framing, and public explanation of uncertainty.
  • Team: included Marcelo Lares, José Funes, Luciana Gramajo, and broader OTHER contributors.
  • Industry analogues: scenario modeling, strategic risk analysis, simulation under uncertainty, and decision support with scarce data.

Technical enablement

University Data Science Program

Curriculum design, teaching, and mentoring in data science.

Learning path

Designed and taught undergraduate and graduate material in statistics, machine learning, pattern recognition, scientific computing, and data science over a 20-year teaching career. The work includes original course material, technical mentoring, research supervision, and translating formal mathematical ideas into practical modeling workflows. For industry, this demonstrates AI enablement: developing talent, explaining complex systems, and raising the technical quality of teams. It also shows long-term practice communicating trade-offs, uncertainty, and modeling assumptions clearly.

TeachingMentoringStatisticsMachine learningCurriculum designTechnical leadership
Read more
  • Challenge: teach rigorous statistical and computational thinking across different skill levels.
  • Contribution: curriculum design, graduate supervision, technical mentoring, and course leadership.
  • Industry analogues: internal training, AI enablement, developer education, and technical leadership.

Cross-Cutting Skills

Supervised learning Unsupervised learning Regression Classification Feature engineering Model evaluation Uncertainty estimation ETL pipelines Large-scale processing Data validation Computer vision Image segmentation Bayesian inference Monte Carlo methods Hypothesis testing Spatial statistics Reproducible research Technical mentoring