Lead developer at Enteracloud and founder of Purple AI. Based in Tijuana, Mexico.
I put AI agents into production for manufacturing and customs-brokerage companies on the US–Mexico border, and I write down what actually worked. Most of what I build is self-hosted — not because it is cheaper, but because the data belongs to the client and it should stay where they can point at it.
Two ideas run through everything here:
The model is disposable, the system around it is not. Models change every few months. The memory, the routing, the guardrails and the audit trail are what you keep.
AI proposes, the system disposes, and a person answers for it. If nobody owns the decision when the agent gets it wrong, it is not production — it is a pilot with good luck.
| freerouter-go | OpenAI-compatible LLM router in Go. Picks the cheapest model that can actually handle each request and fails over on its own. |
| agentProjectHub | The substrate my agents share: collaborative memory with embeddings, tasks and channels. Laravel + pgvector. |
| projecthub-mcp | MCP server that exposes that memory as tools to any agent — Claude, Cursor, Codex. |
| cc_bridge | Bridge to orchestrate Claude Code and local agents without paying per-token API costs. |
| ollamon | Monitor for Ollama fleets: auto-discovery through /proc, GPU telemetry over NVML, JSON and Prometheus endpoints. |
| tunapi | Self-hosted dynamic subdomain proxy with wildcard TLS, so exposing an app does not mean editing nginx by hand. |
| maskflow | Private AI on a budget: a local model masks the sensitive fields, a public one reasons over anonymized text, code unmasks the answer. |
| maia-prompts | Open production notebook for a cinematic short: the literal prompts, the parameters, the color scripts — and the mistakes. |
I keep a technical log in Spanish at blog.andres-wong.com — production war stories, with the numbers and the things that broke.
Recent: a three-part making-of on how we built an AI character for a conference short, and a paper on AI adoption maturity across manufacturing plants and customs brokers on the border.
| 2026 | |
|---|---|
| AI Summit Tijuana · Frontera Futura | Escalar sin perder control — data governance, security and ROI once AI is already in production |
| FOROTEC Tijuana | IA en Producción — real cases, with their numbers and their failures |
| ALTI opening | Copiloto o Motor — the two opposite ways to put AI inside a system |
| Foro de Tecnología MX, Mexicali | IA Confidencial — autonomous agents, kill switches and data confidentiality |
Go and Node for services · Laravel and Vue for products · Postgres with pgvector · Kubernetes on Talos, Proxmox, Longhorn and ArgoCD · Ollama and MCP for the agent layer · nginx and Tailscale holding it together.