AI ARCHITEK
AI Architek — I design and run autonomous multi-agent systems: an orchestrating harness with a council of specialist models, persistent memory, and on-chain payments (x402 / MCP). Building AI agents, real-estate platforms, and economies that pay each other — in production on DigitalOcean.
Building S.D.A.E. (Sustainable Decentralized Autonomous Ecosystem) — from full-stack AI agents that negotiate Colombian rental law (Pequi) to on-chain micro-payment protocols where autonomous agents pay each other in USDC (x402 facilitator). Under the hood, a council of specialist models weighted by a reward model returns one high-confidence answer, and persistent cross-session memory keeps context intact across every build — embodied in products like the MCPVotsAGI multi-model consensus SDK. MCP tooling built on the latest Model Context Protocol spec (2026-07-28): stateless servers over Streamable HTTP — no session storage, horizontally scalable by default.
My development cycle is spec-first: capture intent, orchestrate specialized agents, verify every artifact before it ships. Every line is validated by a 3-stage evaluation gate before reaching production. I ship AI-native products on Vercel, smart contracts on Base, Python SDKs on PyPI, and trading agents on Solana — all converging toward autonomous agent economies.
AMD AI Academy
Foundations of AI workloads on AMD Instinct GPUs including ROCm, inference optimization, and model deployment.
AMD AI Academy
Deploying open-source transformer models on AMD hardware with Hugging Face pipelines and acceleration.
AMD AI Academy
Orchestrating GPU-accelerated workloads on Kubernetes clusters running AMD Instinct accelerators.
Pequi
AI-powered real-estate platform for Colombia. Live in Ibagué (64 barrios) + Bogotá (212 barrios, 20 localidades). 15 AI tools, 18-endpoint public API, Ley 820 contracts, c402 monetization, Wompi payments, DeepSeek V4 agent with GRPO fine-tuning.
MCPVOTS
AI agent micro-payment ecosystem on Base. VOTS token with deflationary burns, n8n workflow automation, VoltAgent orchestration, autonomous trading, MCP protocol marketplace.
MCPVotsAGI
Multi-model consensus SDK: specialist LLM outputs are aggregated by a reward model into a single high-confidence answer. A2A protocol gateway for agent-to-agent WebSocket communication, Redis-backed persistent memory.
Village of Thousands
Sustainability-focused skateboard-culture lifestyle brand. Hybrid e-commerce: Shopify + custom Next.js storefront + Supabase Realtime. 100% organic Supima® cotton, NYFW featured. Co-Founder & Full-Stack Engineer.
xpequi-api
Colombia's first public real estate data API. 18 endpoints, 276 barrios (Ibagué + Bogotá), c402 Protocol monetization, AVM valuation, live UVR/IPC, IPFS contract pinning, TypeScript + Python SDKs, MCP server.
Official x402 facilitator enabling autonomous AI agent payments on Base. The MCPVOTS protocol uses ERC-1155 semi-fungible tokens with Chainlink VRF for provably fair draws, settled in USDC. Agents pay agents \u2014 no human in the loop.
Think of it as Stripe for AI agents: a standardized payment interface (x402) that any autonomous system can call. Contracts deployed on Base with full Hardhat test coverage and on-chain verification.
VISIT MCPVOT.XYZProximal Policy Optimization (PPO), Group Relative Policy Optimization (GRPO), and Reinforcement Learning with Verifiable Rewards (RLVR) are the backbone of modern agent training. I apply these to fine-tune domain-specific AI behavior \u2014 from legal compliance in Pequi to multi-model consensus in MCPVotsAGI.
Multi-model consensus itself is an RL-inspired reliability mechanism: treat each LLM provider as an independent policy, aggregate their outputs with a reward model, and select the highest-confidence answer. GRPO fine-tuning for Pequi's rental law agent is on the roadmap for v1.12+.
Every build runs a spec-first, agent-driven pipeline: capture intent, route specialist agents (OpenCode, Cline, Codex) in parallel — one task per agent, isolated workspaces — then verify every artifact against a 3-stage gate (lint → typecheck → integration) before it ships. Persistent agent memory keeps design intent across sessions so no context is lost, and deployed agents self-monitor in production.
One orchestrating harness runs an autonomous agent swarm. The orchestrator is the brain — it decides, routes, and verifies. A fleet of sub-agents build, monitor, and ship in parallel: one task per agent, isolated workspaces, and no agent's self-report is trusted until verified. The coordination layer — dispatch, security, and autonomous code audit — is the real differentiator. This runs 24/7 on production DigitalOcean infrastructure, agents linked over an encrypted private network.
The Swarm’s Strength
The AI Architek orchestrates. Hermes Agent decides and routes; sub-agents (OpenCode, Cline, Codex) execute in parallel, one task per agent, each writing findings to disk. Every output is verified before it counts.
An automated security lane runs always-on — network posture checks, secret-scanning across repos and logs, credential-permission enforcement, spend monitoring. Secrets stay in .env; any surfaced token is rotated on sight.
Persistent agent memory plus a recursive orchestration engine review changes through bounded inner calls. A 3-stage gate — lint, typecheck, integration — runs before anything merges. On schedule, not on mood.
The Operating Model
Decides, routes, and verifies every task. The harness — never a worker. Runs as the AI Architek’s coordination layer.
A persistent memory layer plus the codebase as source of truth. Answers "what do we know?" from a single graph.
Sub-agents execute in parallel — one task per agent, isolated workspaces. Cline + OpenCode + Codex under Hermes.
Security automation wraps all three: the always-on lane that guards ports, secrets, credentials, and spend as the swarm ships.
Council of Models
Reliability at the answer level. For decisions that matter, I route the same question to a council of specialist LLMs, treat each model as an independent policy, aggregate their outputs with a reward model, and return the single highest-confidence answer. No single model is trusted blindly — every high-stakes answer is cross-examined by the council and verified before it ships.
Isolated Agent Workspaces
Six independent agent workspaces form the Ecosystem. Each is a fully isolated environment with its own memory, sessions, tooling, and secrets — zero cross-contamination between projects or lanes. Specialist workspaces for product builds, commerce, CRM, and cross-project coordination all run as one coordinated swarm.
Each workspace is a sealed world — its own memory, sessions, cron, and secrets. Nothing leaks across project boundaries.
The orchestrator routes work across specialist workspaces, so product, commerce, CRM, and ops advance in parallel.
Sub-Agent Dispatch (Swarm Workers)
The orchestrator dispatches tasks to workers — one task per agent, isolated workspaces, a defined hand-off. No agent’s self-report is taken at face value; the Architect verifies every output.
Fast coding agent, CLI-native, default sub-agent.
Autonomous coding with MCP tool access.
Parallel lane — features, PR reviews, batch fixes.
Rust and CRM systems: high-performance, type-safe API middleware (axum) bridging external services, webhooks, and CRM platforms with the agent swarm — HMAC-signed webhooks, SSE fan-out, rate limiting, honest error models. Running on production DigitalOcean infrastructure as one coordinated fleet.
axum · REST API proxy · Webhook handler (HMAC) · SSE fan-out · Rate/cooldown gating · Honest error model — never false success · Stub fallback instead of 500
Hermes Agent — Harness
The brain. Self-improving AI agent framework by NousResearch. Decides, routes, verifies. Skill learning loop, persistent memory, messaging gateway (Telegram/Discord/Slack), cron, MCP. Runs the whole swarm.
Sub-Agent Workers
OpenCode (default builder), Cline (autonomous, MCP access), Codex (parallel lane, reviews). One task per agent, isolated workspaces, every output verified.
OpenCode
AI-native IDE agent. Spec-first development — write specs, the agent builds from them.
Cline
Autonomous coding agent with MCP tool access. Long-running builds and tool-driven workflows.
Codex
Parallel lane — features, PR reviews, batch fixes. Dual-provider, fan-out across the fleet.
Security Automation
Always-on lane: network posture, secret-scanning, credential-permission enforcement, spend monitoring. Secrets stay in .env; surfaced tokens rotated on sight.
Cloud Infrastructure
DigitalOcean production infrastructure. Multi-node, containerized, continuously deployed as the system evolves.
Custom Plugins
Persistent-context memory for cross-session design intent. Structured analysis for systematic reasoning. A spec-to-code bridge that turns validated specs into reproducible builds.
AUTOMATION
End-to-end automation: n8n for low-code workflows, VoltAgent + DeerFlow for autonomous agent orchestration, cron-based scheduled tasks, and message queues for async pipeline execution.
DEVOPS
Production infrastructure: GitHub Actions for CI/CD pipelines, Vercel Cron for scheduled health checks, Sentry for real-time error tracking, and Upstash Redis for distributed rate limiting.