Comment by AlexisAguirre

15 hours ago

Location: Córdoba, Argentina Remote: Yes, remote only Willing to relocate: No Technologies: TypeScript, React, Next.js, Node, GraphQL with Apollo, Playwright, Python, SQLite, Postgres, LLM agents, MCP servers, RAG, promptfoo evals, OpenTelemetry Résumé/CV: https://portfolio-aguirre-alexis.vercel.app Email: aguirrealexis.cba@gmail.com

I do not trust what I have not verified, including my own code.

Full-stack dev, 5+ years, last year building LLM products rather than demos. Everything below is something I found by auditing my own work, not by reading about it.

intent-gate — https://www.npmjs.com/package/intent-gate A job posting with hidden instructions got my CV pipeline to write ten years of Kubernetes into my resume. Technology I have never touched. The system prompt told it to use only what was in my facts file and it obeyed the posting instead. What stopped it was a deterministic validator comparing every named technology against ground truth, no model in the loop. A prompt is a request. A check is a guarantee. I published the routing pattern as a library.

job-hunter — https://github.com/ale-aguirre/claude-job-hunter Autonomous job search agent running on my own search. It taught me that my own verifier was lying to me. It counted a URL redirect as a successful application. Aggregate confirmation sat at 78 percent and looked healthy, until I graded per channel and found four integrations at 100 percent and one at zero out of sixty, silently dead for months. The average was hiding it. That number was on my CV and I removed it.

DocUnify — production, private Document comparison SaaS running at an Argentine auto parts manufacturer, 72 employees, in a domain audited against IATF 16949 and ISO 9001. Two people there use it for their actual work. The part I would defend under questioning is the semantic reclassification after the LLM call, because the model kept flagging paraphrases as real differences and embeddings catch that.

This week — https://github.com/theam/facility/pull/242 Merged a fix into an open source AI SDLC project and filed two issues, one about an agent being able to write the field that marks its own check as platform verified. I reproduced the bug on my own machine first and my first repro was a false negative, because bash was eating the backslashes in my test payload.