Skip to main content

Engineering portfolio

AI & Data Engineering

I implement AI-assisted workflows and statistical data products with explicit contracts between source records, computation and release. The cases below show where models operate, how outputs are checked and who retains authority.

AI & Agent Systems

I keep generated drafts separate from durable records, make publication conditional on explicit approval, and restrict agent context to named, read-only resources.

Media Monitor

An end-to-end governed news workflow combining source ingestion, versioned contracts, editorial model integration, deterministic tests, revision branches and an explicit human publication gate.

Evidence: implementation, tests and public health contract.

Matías Context MCP

A local, read-only resource gateway with bounded UTF-8/JSON reads, startup preflight, named sources, provenance hashes and protection against path and symlink escapes.

Local implementation · not a hosted agent platform

Supporting pattern: Knowledge / Data Contracts, a public architecture reference rather than a flagship implementation.

Data, Measurement & Decision Systems

My data work combines survey methodology, statistical inference, reproducible pipelines and geospatial delivery, with explicit checks on uncertainty and model support.

Accounting Workflows

A Python pipeline for source ingestion, canonical materialization, semantic marts, governed reports, validation and exact-run replay. Public explanation uses sanitized evidence and preserves a privacy boundary.

Implemented pipeline · no client or certification claim

Argentina Poverty Measurement & Atlas

A governed survey-to-census statistical-geospatial pipeline with semantic alignment, model-support checks, explicit estimands and release boundaries between scientific production and public presentation.

Research-grade infrastructure · release scope applies

Engineering methods

I start from the data model, the decision that a system must support, and the boundaries between sources, computation and authority. I favor simple, inspectable interfaces; deterministic processing where possible; bounded model or agent behavior where needed; and explicit tests, provenance and recovery paths. I use coding agents as collaborators in scoped development workflows, while retaining responsibility for design, review and release.

Additional technical evidence

  1. I reconstructed dwelling, household and person records from Argentina’s 2022 Census RedatamX files and verified the relationships in one census area containing 6,992 people. This is a bounded result, not a completed national extraction.

    Open public extractor documentation
  2. I built a compiler that turns a validated operating snapshot into typed work items, bounded execution packets and traceable run receipts.

    Inspect Office Auto Lab
  3. I directed a Licenciatura en Ciencias de Datos thesis at FCEN-UBA comparing supervised-learning models on EPH household-survey income data.

  4. For micro-to-macro volatility accounting, I coded exact aggregate-variance identities in Python and tested signed covariances, inconsistent components and input-order invariance.

    Open the proof case