I build the systems that work while you don't.

Applied AI reliability — production RAG, agents, evaluation, deterministic scaffolding.

Last updated: 2026-08-22. The shipped/high-signal repos below are the current proof; the full catalog lives on the portfolio site.

Most AI demos look great until they meet real users, messy data, or production constraints. I build the opposite.

This isn't cost arbitrage. It's engineering arbitrage.

Now: Founding Engineer at Zap Viagens — privacy-first, self-hosted AI products (CRM automation, campaign generation) on open-source models for service businesses.

Systems that work in production. Not demos that break on real data.

Production AI is the execution surface: RAG pipelines with grounded retrieval, evaluation flows that catch regressions, agent infrastructure that does not trust the LLM, and observability so you know what is failing and why. The recurring move: discover the method with expensive intelligence, encode it as deterministic software and policy, let cheap models execute it reliably. The nondeterminism doesn't disappear — it stops mattering.

Three and a half years (2022–2025) creating and leading the AI function inside a real business — not side projects. Internal assistants, content-generation pipelines, scraping workflows, RAG systems, business automation. The result: 90% content-cost reduction, 50% faster content production.

The interesting part: Expat Money helps people structure their lives across jurisdictions — flag theory, expat planning, financial sovereignty. The AI wasn't decoration. It was how a business built around individual autonomy automated the labor of producing expert knowledge at scale. That's what grounded retrieval looks like in a domain where wrong answers have real consequences.

These are the existing repos I would inspect first.

• WCP-Compliance-Agent-V5 — Production-ready WH-347 compliance platform: deterministic validation decides, the LLM explains, and every decision is traceable.
• aria-agent — Canonical clean-break agent harness: approval gates, execution tracing, and progressive-disclosure Agent Skills support consolidated from mini-agent.
• groundtruth — Canonical production RAG platform: hybrid search, citations, refusal logic — ingestion and deployment topology consolidated from document-intelligence-pipeline and knowledgeops.
• llm-gateway — LLM proxy with routing, guardrails, cost control, and fallback.
• agenttrace — Agent observability and replay SDK with cost attribution and prompt-cost reporting — consolidated from llm-cost-latency-monitor.
• evalforge — Regression testing for RAG and agentic AI: retrieval and conversational evaluations — consolidated from rag-evaluation-lab and ai-support-simulator.
• github-issue-pr-agent — Autonomous issue-to-PR workflow with a safety boundary.

Also writing: Taste Is All You Need — a long-form analysis of agent-harness economics and why scaffolding beats model spend.

The full public repo catalog lives on the portfolio, each project with its own page:

fishraposo.github.io/projects.html

The catalog separates:

• highlighted shipped/relevant systems
• supporting infrastructure
• meta/profile repos
• archive/study repos

Send me your broken AI workflow. I'll tell you what's breaking, why it's breaking, and the fastest path to a system your team can actually depend on.

Canonical surfaces: GitHub profile and the portfolio site.