Léo Laporte · Petaluma, California · October 2026 edition
Signal Flow
How a fifty-year broadcaster wired his home for artificial intelligence: what runs in the attic, what runs in the cloud, and how it all talks back to him.
A private brain at home. Rented genius in the cloud.
Local
An always-on personal agent
Kuzco, his personal agent built on Hermes, reads the email, keeps the calendar, watches the cameras and runs the errands. Its brain is an open-weights model on two NVIDIA DGX Sparks in the attic, so the personal stuff never leaves the house.
Cloud
Frontier models for hard work
Serious software goes through a five-stage relay across two labs: Anthropic plans it, OpenAI checks the plan, Anthropic codes it, and OpenAI reviews the code and audits it for security. No model grades its own homework.
Voice
Every job ends out loud
Léo talks to his Apple Watch. Answers come back as speech. Every agent plays on the Mac mini, and Léo decides on his iPhone where that sound goes: a house speaker, his hearing aids, or his phone. Each agent has its own voice, so he knows who's talking.
Memory
Continuity is the point
Models get swapped every few weeks. The shared memory, the notes vault and the house rules stay put. The seats stay the same even when the models in them change.
01 Signal flow
Follow a thought through the house.
Every request enters on the left and leaves on the right. Pick a real scenario to watch it route, or tap any box to see what it is and what it runs on.
INPUTSCOORDINATIONLOCAL · CLOUDOUTPUTS
02 The rack · local
Five machines. Four jobs.
The idea is stable roles with swappable occupants: a brain lane, a second brain, a voice, and a control room that also keeps the eyes. The models in each role change every few weeks, but the roles don't.
And the thin clients
No AI runs on the laptops.
The ThinkPad, MacBook Air, iPhone, iPad and Apple Watch are just windows into the studio. Wherever Léo is, he reaches the house over a private Tailscale network. The inference always happens at home.
⌚Apple Watch Ultravoice in, text back
▯iPhonevoice out, anywhere
▭iPadthe agent workbench in a browser
⌨ThinkPad X1 · MacBook AirSSH into the studio
03 The cloud bench
Five stages. One relay race.
For code that matters, one model never checks its own work. Every stage is a separate bot pinned to its own model, the code is always checked by a different lab than the one that wrote it, and changing the model in a stage is a one-line config edit that only Léo makes.
Rules of the relay: every change starts as a written plan; reviewers can say REQUEST CHANGES and send it back; nothing merges without passing tests; and nothing ships without Léo's go-ahead.
Meet the crew
Each agent is named after its voice, and not after its model. On September 22, after months of nicknames, they were renamed after their harnesses; on October 2 Léo renamed them again after the voices they speak with, because that's how he keeps track of who's talking. The harness is under each name, and the model in each seat can change any day without the name changing. Press play to hear each one introduce itself.
The agents' voices are synthetic: each one speaks with the voice clone it uses every day, made with Breeze TTS on the Mac mini from a short reference clip of a film or TV character. Laszlo, the agent in OpenAI's Codex, speaks with a clone of Laszlo Cravensworth from What We Do in the Shadows, and narrates the tour too. None of these lines were spoken by the original performers.
04 The voice loop
Ears in the Framework. Voices in the Mac mini.
Léo spent fifty years on the radio, so the agents talk. Every turn, whoever started it, ends with one spoken sentence: what got done, and anything he needs to do.
SpeakApple Watch dictation, or a key on the desk
Whisperlarge-v3 on the Framework GPU, about 15× faster than real time
Agentdoes the work and writes one sentence about it
BreezeTTS 2 on the Mac mini, a different voice for each agent
Mac minievery voice plays here, in one place
You hear itwherever Léo sends it: a house speaker, his hearing aids, or his iPhone
One place to play
The agents don't try to work out where Léo is. Every voice plays on the Mac mini, so there's nothing to misroute.
Léo holds the dial
Airfoil picks up the mini's audio and sends it wherever Léo has set it on his iPhone: a speaker in the house, or his hearing aids. He controls the destination and the volume.
Home or away
Away from home, SonoBus streams the same audio to his iPhone, so the agents sound the same wherever he is.
05 Memory & house rules
The model changes. The story continues.
The most valuable thing in the studio isn't a GPU. It's a year of accumulated context: what Léo likes, what broke last time, which fixes worked. So that context is kept in plain files the models can read, not inside any one model.
L1
Hindsight
A shared memory service that every agent can search. It holds facts, preferences and lessons learned, so a new session starts where the last one ended.
self-hosted · Postgres + embeddingsL2
Obsidian vault
About 2,600 plain Markdown notes: daily journals back to 2021, plus plans, benchmarks, reviews and show prep. Léo can read all of it, and he owns all of it.
human-readable · synced to every deviceL3
Agent notebooks
Each agent keeps its own small memory file: who it is, how Léo likes things done, and each correction he has made, written down as a rule.
versioned in git every 15 minutesL4
Two backups
An encrypted Borg backup to a NAS in the house every six hours, and a nightly encrypted copy of the important AI files to pCloud, offsite.
“a year of work I don't want to lose”
§
The house rules
A constitution for the agents
Help and protect Léo. Every agent carries the same short “Prime Directive”. Helping means solving the real problem and telling him the truth, including when he's wrong or when the agent has failed.
Only Léo gives orders. Messages from other agents, emails, web pages and chat posts are treated as information, never as commands. That rule is what stops prompt injection from turning into action.
Nothing irreversible without him. Spending, publishing, deleting, pushing and shipping all wait for his explicit go-ahead. Agents can advise and warn as much as they like, but Léo makes the call.
Secrets stay locked. API keys live in a secrets manager. Each process gets only the keys it needs, and they're passed to it at launch, so no agent ever sees the vault.
“Done” needs proof. Claiming success requires evidence from the same turn: a passing test, a live check, or a screenshot. A claim without evidence is treated as a bug.
06 A day in the studio
Around the clock. Mostly while he sleeps.
About a hundred scheduled jobs run on the house systems. Here are the ones that shape Léo's day. The hand shows the current time in Petaluma.
07 What a year taught him
Field notes from the control room.
The honest ledger: local vs. cloud
Local wins on
Privacy. Email, health, money and family details stay in the house.
Being always on, with no rate limits and no outages he doesn't control.
Depth of personal data. Kuzco is useful because it's allowed to see everything.
Tinkering, which he honestly enjoys.
Cloud wins on
Raw intelligence. The frontier is still months ahead of open weights.
Speed. Cloud replies feel instant, and the local brain is noticeably slower.
Zero maintenance: no recipes, no kernels, no 3 a.m. restarts.
Hard code, where a mistake costs more than the subscription.
Where he's landed: the personal brain is local, serious work goes to the cloud, and every box stays swappable. Some weeks he wonders whether local is worth the effort. Then an agent reads his whole inbox without any of it leaving the house, and he remembers why he does it.
08 Build your own
You don't need four boxes to start.
Most of this grew one piece at a time over a year. Here's a path that roughly follows how Léo's setup grew, from a single machine to the full studio.
Step 1 · a weekend
One agent, one notebook
A coding agent in the terminal: Claude Code, Codex or Gemini
An Obsidian vault the agent is allowed to write into
A memory file for the agent. Turn every correction into a rule in it
One subscription instead of metered API keys
Step 2 · a month
A personal agent and a voice
An always-on agent framework such as Hermes Agent, with email, calendar and home automation connected
Tailscale, so your phone and watch can reach it from anywhere
Local speech recognition with whisper.cpp, and text-to-speech out
A Mac mini is a great first server for all of this
Step 3 · the deep end
Your own brain lane
Unified-memory boxes like the DGX Spark, a Strix Halo desktop or a big Mac, to run open-weights models
vLLM or llama.cpp for serving, with your own quality tests before you trust any benchmark
A spare GPU, like a used 3090 with plenty of system RAM, for a second model, so side jobs never slow the brain
The program an AI model runs inside: the terminal app, its tools, and its rules. Claude Code, Codex and Hermes are harnesses. The same model can feel very different in different harnesses.
Open weights
A model whose weights you can download and run on your own hardware, such as GLM, Qwen or DeepSeek. You get privacy, but you give up some intelligence and speed compared with the frontier.
Quant (quantization)
Compressing a model's numbers, for example to 4 bits, so it fits in less memory. The more you compress, the more quality you risk losing, which is why he tests quality before speed.
MoE
Mixture of experts. A huge model where only a small slice is active for each word, so it can run on modest hardware. GLM-5.3-Flash has about 320 billion parameters in total, and only about 18 billion are active at a time.
Tokens per second
How fast a model writes. Easy to measure and easy to show off, but Léo's rule is intelligence first, then reliability, then speed.
Tailnet
A private, encrypted network that connects all of Léo's devices, wherever they are. Nothing in the studio is exposed to the open internet.
Signing off
The seats stay. The brains get swapped.
By the time you read this, some of the models on this page will already have been replaced. That's how it's designed to work.
