Skip to content

The Lab

Where it breaks before it touches a patient or a student.

Consolidate, don't expand · squeeze what already exists

In short: a small fleet serves real conversational agents and tools, in production. Inference runs mostly on owned hardware —about 99.5% of requests over the 30 days measured in July 2026— and escalates to cloud models through a router with automatic fallback.

Irreversible actions go through code-level gates, secrets live in an encrypted vault, and backups follow a 3-2-1 scheme. Anything new —models, settings, agents— is first measured in my own arenas with blinded judges and only reaches production with data.

Updated October 7, 2026

Download PDF

Agents

1:1 support agent

WhatsApp · answers health-companion users, opens tickets, learns from every conversation.

Read the note

Group agent

WhatsApp · its own identity and model per group, keeps context, encourages participation.

Read the note

Collections agent

Telegram · payment follow-up wired to the CRM.

Multimodal agent

Voice, text and image, over Telegram and web.

Prototyping bench

Where new things are tried before they reach production; currently home to the multimodal agent.

Personal assistant

Telegram and web · long-term memory, email digests, reminders and tasks.

Read the note

Raspberry Pi 5 · 16 GB — Mac mini M1 · 8 GB (prototyping bench)

Brain / Orchestration

Orchestration station

It coordinates; it doesn't do the heavy lifting.

Coding-agent orchestrator

Claude Code, with a second opinion from Codex and Antigravity CLI.

Local speech and text models

MLX.

Secrets and single sign-on

Encrypted vault with rotation + single sign-on · Vault + Keycloak.

Mac mini M4 Pro · 64 GB

Services

Internal clinical hub

Cancer center · internal tools, access per user.

SimAcademy

Moodle-based medical education platform · courses and certificates.

Read the note

Asesores en Emergencias

Clinical training platform.

AI-assisted marketing studio

In production

An agent with its own interface: images, brand voice and publishing for the in-house brands. Runs on the same machine as the clinical training platform.

Read the note

Browser arm (Explorer)

Shared by the orchestrator and the personal assistant.

LLM router

One door to every model; switching models doesn't mean rewriting agents.

Inference

Local inference with vision

A 35B MoE Qwen model that reads images, served locally with vLLM, at no cost per use.

Read the note

Irreversible-action classifier

In production

Stops destructive actions before they run.

Cloud models

Anthropic, OpenAI, Gemini and others, always through the router.

Model evaluation arenas

I pit models against each other on my own cases and criteria —task success, tool-use safety, calibration, latency— before anything reaches production.

Read the note

AI workstation · 128 GB unified memory

Storage and backup

Primary storage

9 TB · fleet library and private file cloud (Nextcloud).

Offsite backup

Encrypted copy off site · borg · 3-2-1 scheme.

Network

Encrypted private network

Across every node · Tailscale.

Three paths to the internet

Fiber · cellular SIM · satellite, with automatic failover.

Monitoring

  • Per-service watchdogs that restart and report.
  • Twice-weekly fleet review: it detects and proposes; nothing runs without authorization.
  • Phone alerts through the personal assistant.

Ground rules

  • Consolidate, don't expand

    Squeeze what exists before adding hardware or complexity.

  • A gate for the irreversible

    Destructive actions are stopped in code, not in the prompt.

  • Secrets only in the vault

    Never in repos or loose files; rotated and distributed.

  • Nothing without written authorization

    Anything touching production is approved first.

See all technical notes