Miguel Sanchez Senior Software Engineer

Regulated‑software rigor, applied to AI‑written code.

I evaluate code written by AI coding agents, looking for the failures that matter in production: concurrency bugs, broken data integrity, and unhandled edge cases.

How I evaluate

  • Oracles come from requirements, not from the code.

    Expected results are derived from specifications and boundary conditions, never by reading the implementation under test.

  • Passing tests prove only what they exercise.

    Every result states what was covered and what was not.

  • Two models agreeing is not evidence.

    Correlated AI reviews count as one weak signal, never as independent validation.

  • Findings are reproduced before they are reported.

    Each defect comes with an executable test, and each finding is challenged before it is published.

Case studies · In progress

  • Evaluating Concurrent Code in .NET

    Race conditions, deadlocks, and UI-thread violations in agent-written C#.

  • Evaluating Python Code for Data Validation

    How agent-written validators handle malformed, boundary, and adversarial input.

  • Validating LLM Outputs in Python

    Enforcing structure and correctness when an LLM sits inside a data pipeline.

First case study in preparation.

Background

15+ years building software. I work on FDA-regulated medical device software (IEC 62304, ISO 14971), focusing on code correctness, architecture, and product cybersecurity. I also taught Software Engineering and Network Programming at Universidad ORT Uruguay for a decade.