Ricardo A.

Most AI demos look great in a Loom and break in production. The agent loops on a malformed tool response. Token bills triple in a week. Retrieval returns nothing for the question that mattered. The model hallucinates a refund policy that doesn't exist. Nobody saw it coming because nobody built the harness underneath. I'm Ricardo Argana. I spent 4,400 hours testing production systems before I started building AI ones. I know exactly what fails at 2am because I'm the guy who used to find it. That's why my Claude API integrations, MCP servers, RAG pipelines, and multi-agent systems ship with auth, error handling, observability, and tests. Not vibes. What I build Custom MCP servers connecting Claude, GPT, and Gemini to your databases, APIs, and internal tools. Multi-agent systems with orchestration, tool use, structured outputs, and real failure recovery. RAG pipelines with eval sets, re-ranking, and full retrieval logs so bad answers are debuggable. Production Claude API integrations wired into Python/FastAPI and NextJS backends. Full-stack AI platforms on PostgreSQL, Supabase, AWS, Vercel, Hetzner with WebSocket, Socket IO, and Redis pub/sub. Recent work, with numbers Remake AI. Cloud robotics platform across 10 interconnected repositories. 4,200+ line Express and Socket IO backend. Three-phase HMAC-SHA256 handshake. 12,000+ line ROS2 package on physical robots with 106 tests. Sub-second latency, browser to cloud to physical robot. Physical MCP. Open-source AI agent safety layer for real robots. 100 tools, 2,167 tests, published to npm at v1.0.0. Geofencing, velocity limits, e-stops, deadman switch, audit logging. MCP Operator. Self-hosted no-code AI agent platform on Next.js and Supabase. Cron triggers, HMAC-verified webhooks, self-correcting feedback loop where each run sees the last 5 results. AI Dev Agency. 5 Claude-powered agents (PM, Designer, Frontend, Backend, QA) on a message bus orchestrator. Real code written to disk. Amplify MCP + Hetzner MCP. Shipped a Hetzner production app inside 48 hours on a platform I had never touched, by ingesting all the docs into Claude-callable tools. GFY. Agricultural platform on PostgreSQL with SQL pipelines and reporting dashboards. Live in production. How I work You send the brief. I send a scoped plan back the same day. Weekly milestones with daily async updates and Loom walkthroughs. Start small so you can evaluate the work before committing to scope. For advisory engagements, the first session is 90 minutes fixed-price with a written one-pager back within 48 hours covering architecture, eval design, system prompt scaffold, and the failure-mode hierarchy. Then we decide if we keep going. Capacity 30+ hours per week, US/EU timezone overlap from the Philippines. Currently rebuilding Top Rated status, which means I take serious work at fair rates and prioritize 5-star outcomes on every contract. If you need an AI engineer who can ship Claude API integrations, MCP servers, RAG pipelines, or multi-agent systems that survive production traffic, send me a message. No commitment, just a real conversation about your specific case. Honest work, for the glory of God. - Ricardo Argana