Physics student (UBA) · Applied AI builder · AI agents, evaluation & verification · Scientific computing for energy
-blue?style=flat){kind=icon}
Advanced Physics student at the University of Buenos Aires (expected 2027), oriented to probability, statistics and mathematical modeling.
I build software and AI systems by orchestrating AI agents. The agents write most of the code; I keep the decisions that define the quality of the result: the specification, the architecture and the validation criteria. As a physicist, what I care about is that the result is correct and measurable, not just that it runs.
- Spec-driven development with AI agents: a Claude Code harness where every change starts as a spec with numbered acceptance criteria, approval is stored as a hash, acceptance tests are written first, and deterministic CI gates decide when work is done.
- Reliable AI: retrieval-augmented systems that cite their sources or refuse to answer, evaluated against a human-curated golden set with statistical comparison of configurations.
- Hardware & control: at TECSCI I designed a nanometer-precision linear stage for laser micromachining, including its closed-loop control firmware, and raised the encoder resolution by a factor of 60 by rearranging its magnets.
- Energy: Python tooling for drilling mechanics, 1D geomechanics and decline curve analysis in Vaca Muerta.
| Project | What it does | Stack |
|---|---|---|
| offshore-incident-rag | Grounded Q&A over ~1,300 public BSEE offshore incident reports and Safety Alerts: every claim cites document and page, or the system declines to answer. Built spec by spec with an SDD harness (hash-bound human approval, AC → test traceability, ruff, mypy strict, evaluation regression gate). In progress: ingestion done, structure-aware chunking under way, hybrid retrieval and LangGraph answering next. | Python PostgreSQL pgvector LangGraph Docker Claude Code |
| Project | Core Domain & Physics | Stack |
|---|---|---|
| vaca-muerta-dca | Decline curve analysis and well economics on 1,590 Vaca Muerta horizontal oil wells from public SESCO data (full history 2006–2026): rates per producing day, robust log-space Arps fits with uncertainty, EUR to the economic limit, NPV, and a 24-month hindcast that measures forecast error. | Python Pandas PyArrow SciPy Pytest |
| drilling-optimization-engine | Post-well drilling analysis on a reproducible synthetic benchmark: unit-aware LAS ingestion, Mechanical Specific Energy (Teale / Dupriest), founder point per formation by hinge regression with bootstrap intervals, a data-driven WOB road map, and stick-slip analysis on time windows. 70 tests in CI. | Python Lasio SciPy NumPy Pytest |
| geomech-fracability-engine | 1D Mechanical Earth Model on a synthetic Vaca Muerta pilot well: overburden, Eaton pore pressure, poroelastic horizontal stresses, landing selection and a 2,500 m lateral with stages partitioned to minimise Shmin contrast, plus QC masks for every clipped value. 49 tests in CI. | Python Lasio NumPy Matplotlib Pytest |
| Project | Description | Stack |
|---|---|---|
| uMachining | Closed-loop motion control firmware for a linear laser micromachining axis on a dual-core RP2040: cascaded PID, trapezoidal motion profiles, least-squares encoder calibration, homing and a G-code parser. | C/C++ RP2040 PlatformIO SPI PID |
- AI-assisted engineering:
Claude Code(subagents, skills, hooks), spec-driven development, human approval at key decisions - Verification & evaluation: test-first acceptance tests, deterministic CI gates, retrieval evaluation (recall@k, MRR, faithfulness), bootstrap confidence intervals and paired tests
- Scientific computing & data:
Python,NumPy,SciPy,Pandas,PyArrow,Scikit-Learn,Jupyter, non-linear fitting, numerical simulation - Infrastructure:
Git,GitHub Actions,Docker,PostgreSQL+pgvector - Embedded:
C/C++written with AI agents and validated on hardware,RP2040,SPI, closed-loop control - Domain modeling: decline curve analysis, drilling mechanics (MSE), 1D geomechanics, metrology & sensor calibration