Arthur S.

I build complete products end to end, solo: PostgreSQL schema, typed API (tRPC), React or React Native frontend, AWS infrastructure as code, keyless CI/CD, custom domain, and TLS. 6+ live products run on a serverless platform I designed and reuse, and my recurring infra bill trends toward $0 by deliberate cost engineering: a documented S3-outbox relay pattern that replaces NAT gateways, free-tier tuning, and Durable Object hibernation. I am also an AI/LLM engineer in the strong sense: multimodal apps (vision + voice + RAG with pgvector), multi-provider orchestration (Anthropic, OpenAI, Gemini), LangGraph agents, LLM-as-judge evals, and an open-source agent framework I built on the Claude Agent SDK + Model Context Protocol: 20 agents, 47 skills, and 10 MCP tools that autonomously design, implement, and verify code behind human gates. Don’t trust the portfolio. Test me. Two ways: 1. AI briefing pack. Ask me for it, paste it into your own ChatGPT or Claude, and let YOUR AI evaluate my credentials. It is distilled from an automated audit of 26 of my real repositories and includes what I am still improving, because an honest document is more useful to you than a polished one. 2. Prompt challenge, free, one round. Send me a realistic prompt from your project: feature spec, bug report, or architecture question. I return the output of my AI-assisted workflow plus my human review: what I kept, what I corrected, and why. You judge the result and the judgment. Stack keywords: TypeScript, React 18/19, Next.js, Expo/React Native, tRPC, Drizzle, PostgreSQL, Supabase, AWS (Lambda, SAM, API Gateway, RDS, DynamoDB, CloudFront, Cognito), Cloudflare (Workers, Durable Objects, Pages), GitHub Actions OIDC, Clerk, LangGraph, Claude Agent SDK, MCP, RAG, Godot 4, n8n, Make, Zapier. Portfolio item blurbs Xenodot Forge (open source, MIT) = AI game-dev framework on the Claude Agent SDK + MCP: a human-gated design → implement → verify agent pipeline (20 agents / 47 skills / 10 MCP tools, ~14.7k LOC) driven from a WebSocket UI, with a verification-first quality gate (headless engine checks, zero-warning lint, CI-enforced validators). Used it to build a playable cross-platform 3D FPS in Godot 4.6. Available on my GitHub. Probius.app (live EdTech) = Study platform for the Brazilian OAB bar exam. React + Vite SPA on S3/CloudFront, Lambda + tRPC + Drizzle + PostgreSQL, NAT-less VPC posture with all outbound AI/email traffic via an S3-outbox relay, deployed exclusively through keyless GitHub Actions OIDC. Multimodal AI mobile app (~28k LOC) (3itch ai) = Fashion-critic app: Expo client, tRPC-on-Lambda backend, Cloudflare Durable Objects realtime engine. Vision (Gemini), persona brain (GPT-4o), voice in/out (Whisper + TTS), wardrobe RAG (pgvector), per-user AI spend caps, offline RS256 JWT verification on both backends, and gitleaks in CI. Production e-commerce (~31k LOC) = Luxury-furniture platform: Next.js 16 SSG storefront, React admin SPA, serverless API, and one PostgreSQL backend. Infra cost cut to ~$0 with the S3-outbox relay pattern, plus a single-query JSONB product search function that eliminates N+1 fetches. Multi-tenant PMS + OTA sync = Vacation-rental management system with Channex integration to Airbnb and Booking: idempotent booking sync, SQS worker queue, EventBridge polling, and a 12-scenario automated certification suite. Prompt Challenge I will run your prompt through my AI engineering workflow and hand you the reviewed output. Description: Send me one realistic prompt from your project: a feature spec, bug report, refactor, or architecture question. I run it through my AI-assisted workflow, using the same agent tooling I open-sourced, and deliver: (1) the raw output, (2) my line-by-line human review marking what I kept, what I corrected, and why, and (3) a short note on how I would productionize it. This shows you what you are actually hiring in 2026: not “can he prompt an AI,” but “can he judge, correct, and ship what the AI produces.” First round free on invite.