Senior AI Fullstack Engineer in Lisbon, remote worldwide. 14 years shipping — 4+ deep in web3, 2+ building AI developer tools. I build AI-native products, agent tooling, and the harnesses that keep them honest: bounded specs, isolated worktrees, evals that fail closed.
My portfolio speaks MCP. Point your agent at https://www.rubenmarcus.dev/api/mcp and let it read this résumé instead of you — setup below.
| Project | What it is |
|---|---|
| Ralph Starter | Open-source AI coding orchestrator. Swarm mode — race / consensus / pipeline over isolated git worktrees — plus an MCP server and a Figma→code visual validation pipeline. |
| Autoresearcher | Benchmark-driven autonomous research CLI. Divergent agent populations in isolated worktrees, champion merging, a Pareto frontier of candidates, keep/reject validity gating. |
| AEO.js | Answer Engine Optimization framework. AI-crawler policy analysis and LLM-ready exports (llms.txt, ai-index.json). Astro + Next plugins. |
| AEO Checker | The hosted scanner on top of it — 4,569 scans across 2,259 unique sites. |
| ECDSA.fail — #1 | Led AI engineering of an autonomous multi-agent research harness: 9 specialist LLM roles, 7+ providers, role→model routing, fail-closed adapters, spend gates. Research contributor on the resulting publication. |
| QEC Decoder — #1 | Top of Optimization Arena's quantum error-correction leaderboard (2,642 EPM, near Bayes-optimal) via a multi-agent campaign with an anti-overfitting evaluation protocol. |
| CS Brasil | Browser FPS built with an agent gauntlet. WebGL, no install, no netcode. 2,191 players, 154K+ kills, 27 countries in alpha. |
| Mirofi.sh | Open-source multi-agent social-simulation engine (GraphRAG/Zep, OASIS) productized into a hosted SaaS. |
Before that: MultiVM Labs / Quantum (post-quantum L1 — main site, block explorer, and an ML-DSA-65 smart wallet on Arbitrum Stylus), Bitte Protocol (top human committer on the production AI runtime, #1 on the agent SDK monorepo, #3 on the wallet), Grover, Zup Innovation / Itaú Open Banking, Santander, and a long agency tail — Under Armour, Centauro, Samsung, Panasonic, Monsanto. Full archive →
| 2.85M+ | agent messages processed in production (Bitte AI runtime) |
| 24,164 | unique users on agents I built · 16,703 agents deployed on the runtime |
| 32 | AI agents and loop roles built · 25 versioned agent skills |
| 4,569 | AEO scans across 2,259 sites |
| 34K+ | all-time npm downloads across 8 packages |
| #1 · #1 | ECDSA.fail · Optimization Arena QEC decoder |
| 14 years | shipping · ~2M lines of code, career estimate |
No account, no auth, no signup. Four read-only tools plus one that emails me a brief.
https://www.rubenmarcus.dev/api/mcp
| Tool | Returns |
|---|---|
get_resume | full resume: experience, skills, proof points, links |
get_services | the fixed-scope offers |
check_availability | current engagement status |
book_intro | posts a project brief to my inbox |
Setup per client
Claude — Settings → Connectors → Add custom connector, name it rubenmarcus, paste the URL. Or in Claude Code:
claude mcp add --transport http rubenmarcus https://www.rubenmarcus.dev/api/mcpChatGPT — Settings → Apps & Connectors → developer mode → Create, paste the URL, save.
Cursor (~/.cursor/mcp.json), and any client that takes a streamable-HTTP URL:
{
"mcpServers": {
"rubenmarcus": { "url": "https://www.rubenmarcus.dev/api/mcp" }
}
}Kimi — add a streamable-HTTP MCP server with the same URL, then run tools/list.
Codex CLI / any stdio-only client (~/.codex/config.toml) — bridge it:
[mcp_servers.rubenmarcus]
command = "npx"args = ["-y", "mcp-remote", "https://www.rubenmarcus.dev/api/mcp"]Then ask: What has Ruben shipped? · Summarize Ruben's experience with AI agents. · Is Ruben available right now? · Book an intro — I want to build a trading bot.
No MCP client? Plain endpoints.
| Resume JSON | https://www.rubenmarcus.dev/api/resume.json |
| Resume text | https://www.rubenmarcus.dev/api/resume.txt |
| Agent guide | https://www.rubenmarcus.dev/AGENTS.md |
| Connect guide, as Markdown | https://www.rubenmarcus.dev/connect.md |
| LLM index | https://www.rubenmarcus.dev/llms.txt |
| MCP server card | https://www.rubenmarcus.dev/.well-known/mcp/server.json |
| Agent skill | https://www.rubenmarcus.dev/.well-known/agent-skills/portfolio-mcp/SKILL.md |
Every blog post has a .md twin: append .md to any post URL.
Three origins: a production fleet that signs transactions, a research command center, and a game harness. Full directory →
Bitte Protocol — the production fleet
| Agent | Chain | What it does |
|---|---|---|
| AI Framework | EVM · NEAR · Sui · Cardano · Midnight | Natural language in, signed transactions out. The runtime everything below runs on. |
@bitte-ai/chat + make-agent | chain-agnostic | An OpenAPI spec is the agent. |
| Uniswap agent | Ethereum / EVM | Keyless cross-chain swaps from a chat prompt. |
| Gnosis Pilot | Gnosis | A DeFi yield copilot with live data. |
| Polymarket agent | Polygon | A prediction-market analyst that can place the bet. |
| meme.cooking agent | NEAR | One prompt, one memecoin. |
| Solana Assistant | Solana | Solana chain data, agent-ready. |
| Jupiter swap agent | Solana | Quote and swap, one endpoint. |
| Morpho agent | Ethereum | Lending and borrowing, spelled out for an LLM. |
| Aerodrome agent | Base | The full veAERO machine as agent tools (~25 tools). |
| Walrus agent | Sui | Decentralized storage by chat. |
| Sui assistant | Sui | A Sui explorer that talks back. |
| ENS agent | Ethereum | Names, records, registrations as tools. |
ECDSA.fail — the command center that took #1
Frontier Dissector · Circuit Engineer · Density Analyst · CUDA Engineer · Pod Manager · Research Scout · Orchestrator-Reviewer · Combinator — 9 specialist roles routed across 7+ providers, with contracts, spend gates and fail-closed adapters. How it worked →
CS Brasil — the Gauntlet
Three rules hold the loop together: the ruler is not negotiable (not "looks good" — wins or loses against a CS2 frame, by which measure), the builder never grades its own work, and the loop ends when I stop it, not when an agent declares itself satisfied. Every claim has to carry a number and a file:line.
| Role | Contract |
|---|---|
| Measured baseline | Captures every map × 2 aspect ratios × 4 angles before anything changes. No baseline, no A/B — and without A/B the loop is just opinion. |
| Graphics critic | Grades the frame, never the builder's report. Sees pixels and code, nothing else. |
| Map fidelity critic | Grades the map against how the real Brazilian place actually looks. |
| Weapons critic | Visual and feel scored separately. |
| UI critic | Menu and HUD scored separately. |
| Gameplay critic | Movement, bots, combat flow. |
| Parallel builders | Partitioned by a generated file/range conflict table, so several agents edit one large file at once without colliding. |
| A/B verifier | Re-shoots the same frames and proves the delta, or the round didn't happen. |
| Regression Hunter | The agent that pays for the whole loop: a visual win that breaks the game is a regression. |
| Bug Hunter | The ruler comes before the fix, and a mutation has to prove the ruler bites. |
| Régua | Writes the invariant, the probe and the gate — 25 consistency criteria that outrank the fidelity bar. |
| Asset review | Adversarial critic on every new character, map, model or texture before the front can be called done. |
| Content & faction pipelines | Build a team, a real-world map, or a whole faction end to end — roster, crest, cover, 3D character, thumbnail, selection video, original voice. |
Critics run in parallel with clean contexts; builders run in parallel across four checkouts of the repo on separate branches, plus dozens of throwaway git worktrees for one-off fixes.
Versioned method, not loose prompts — 25 skills that encode how the work is actually done. Browse them →
Ralph Starter Orchestration · Autoresearch Benchmark Loop · AEO Delivery System · Benchmark Frontier Archaeology · ECDSA.fail Route Spec · Circuit Engineer Factory · ECDSA Circuit Optimization · Reversible Circuit Validation · Peak Qubit Reduction · Toffoli Reduction · ECDSA.fail Island Hunting · Multi-Agent Research Collaboration · CS Brasil Content Pipeline · Faction Pipeline · Adversarial Asset Review · Régua / Executable Quality Gate · CS Brasil PR Triage · CS Brasil Smoke Check · Gauntlet FPS · Bug Hunt · Blog Voice · Bilingual Publishing · Portfolio Cover System · Frontend Delivery Harness · Portfolio MCP
Every post ships in EN and PT, and every post has a Markdown twin for agents — append .md to the URL. All posts →
Earlier posts
| Date | Post |
|---|---|
| 2026-07-11 | Cross-pollinating LLMs: peer review for machines |
| 2026-07-08 | Keeping an autonomous research agent honest |
| 2026-07-02 | Evals are the product |
| 2026-06-25 | Git worktrees are my agent orchestrator |
| 2026-06-18 | The swarm that took #1 on ECDSA.fail |
| 2026-06-05 | Context engineering inside a runtime with 344K chats |
| 2026-05-28 | How RAG works inside Mirofi.sh |
| 2026-05-12 | The Mini Shai-Hulud Case and the Real Risk of Dependencies |
| 2026-04-16 | How I Hit #1 on a Quantum Error Correction Challenge using AI |
| 2026-02-19 | Automating entire workflows with ralph-starter |
| 2021-05-16 | Getting started with Next.js + Strapi: Security first |
| 2021-05-07 | Why use Next.js + Strapi? |
aeo.js · ralph-starter · autoresearcher · make-agent · scanrepo · elendil · new-agent · qday
- Languages — TypeScript · JavaScript · Rust · PHP · C#
- Frameworks — React · Next.js · Svelte 5 · Astro · Vue · Angular · Node · Bun
- AI & agents — Claude SDK · OpenAI SDK · AI SDK · MCP · Mastra · Ralph loops · custom harnesses
- Web3 — EVM · NEAR · Sui · Solana · Cardano · Wagmi · Viem · ERC-4337
- Infra — AWS · GCP · Vercel · Docker · Terraform · GitHub Actions · Turborepo
- Craft — Figma · Storybook · Tailwind · Three.js · GLSL
Selectively available for full-time roles and freelance contracts. Fixed-scope engagements preferred. I reply within a day or two.
- AI products & agent systems — idea or brittle prototype → product with a harness, evals, observability. 3–8 weeks.
- AI-native frontend & design engineering — a distinctive web product with an AI delivery harness that preserves quality. 2–6 weeks.
- AEO audit & implementation — make a site legible and citable to answer engines. 2–4 weeks.
ruben@rubenmarcus.dev · rubenmarcus.dev/contact · or just let your agent call book_intro.