September 15, 2026

Currently thinking a lot about evaluator-first architectures. If we can't deterministically evaluate a pipeline in CI/CD, we have no business shipping it to production. The next phase of AI engineering is entirely about writing better programmatic judges, not just better prompts.

def evaluate_provenance(claim: str, context: str) -> bool:
    """Deterministic check before LLM evaluation."""
    if not is_substring_fuzzy(claim, context):
        raise HallucinationError("Claim not found in context")
    return True

August 20, 2026

Just finished re-reading The Master Switch. It's fascinating how the current AI landscape mirrors the early days of radio and telecom. The centralizing forces are identical.

July 10, 2026

Open source contribution merged into Ragas today. Fixing metric integrity bugs is unglamorous work, but it's the only way we build trust in these systems.