Enrique Rubio

Blog

Building in public

An engineering journal, not a content calendar: real incidents from Traffic OS and my other systems, the architecture decisions behind them, and what I'd build differently next time.

We build and use our own AI and automation systems.Building with AI →

Architecture··2 min read

Where your private data lives in a local AI setup

A walk through the channels private data can leak through in a local AI setup (Git, bind mounts, builds, logs, worktrees) and how one directory was picked so that none of them reaches it.

Building with AI··2 min read

LUFS correct is not good audio: two gates, one of them a human

An audio bed hit its loudness target and a person still judged it too synthetic. How the code keeps the machine verdict and the human verdict apart.

AI··2 min read

Claude Code: five fixes that protect your work when the agent runs unattended

The latest Claude Code release notes include small changes that matter if you leave it working unwatched: worktrees, binary files, search errors and usage limits.

AI··1 min read

Gemini 4 Argon on AI Gateway: thinking levels instead of budgets

Vercel added Gemini 4 Argon to AI Gateway. What it changes for builders: the model ID, level-based reasoning control, and how pricing is passed through.

Building with AI··2 min read

docker rm -f keeps anonymous volumes: the leak in my Postgres test fixtures

My test fixtures removed ephemeral Postgres containers with docker rm -f and left an anonymous volume behind per container. What the code shows, and the regression test with a negative control.

Architecture··4 min read

My AI System Passed 2 of 3 Tests to Publish Alone. I Decided Not to Turn It On

My blog pipeline passed 2 of 3 autonomous-publishing runs. The third was refused, so it stays on human approval. Why I didn't loosen the approval gates.

Building with AI··2 min read

A LinkedIn timeout is 'unconfirmed', not 'retry'

How Jarvis handles a timeout when posting to LinkedIn: a state of its own for 'I do not know', and a human as the only way to force a retry.

Building with AI··3 min read

From 1.8 GB of Raw Video to a Finished YouTube Video: My Own Editing Pipeline

How my pipeline turned ~1.8 GB of raw video into an 8-minute 1080p YouTube video: upload, audio, motion, metadata, chapters, thumbnails and human review.

AI··2 min read

OpenAI Dots Pricing: What's Included, What Isn't

OpenAI Dots: eligible Pro and Business Premium users get one agent at no extra charge; extra workload has no price yet. Meta's Muse is free for most uses.

AI··3 min read

Gemini 4 Argon: Pricing, Access and Benchmarks

Gemini 4 Argon starts with trusted defenders in Google's Fairwind Program. Launch price: $2/$10 per million tokens, then $4/$20. Benchmarks are self-reported.

AI··3 min read

OpenAI Scraps GPT-6.1 Astra Over Honesty and Authorization Failures, Then Ships Sol a Day Later

OpenAI drops GPT-6.1 Astra because it missed the bar on scope, authorization and honesty with the user. What failed and what it means for agent builders.

AI··3 min read

OpenAI's Mental-Health Benchmark Puts GPT-6 Astra on Top — Scored by Another OpenAI Model

OpenAI's new mental-health benchmark puts GPT-6 Astra on top, ahead of Anthropic's Claude Opus 5.5 and Google's Gemini 2.5 Pro -- graded by another OpenAI model.

AI··4 min read

Google Gives Gemini a Live Avatar Face — and Its Own Safety Report Says the Model Isn't New

Google's Live Avatar streams a lip-synced face onto Gemini's live conversations in 97 languages -- but its own safety report says the model underneath isn't new.

AI··4 min read

GPT-6 Astra: Parallel Halves Its Research Costs While OpenAI Admits It Can Barely Monitor the Model's Reasoning

OpenAI is showcasing Parallel as GPT-6 Astra's success story: half the time, half the cost. The same week, the model's own safety report admits a substantial drop in internal transparency -- and OpenAI's own researchers say they're worried.

AI··4 min read

Ringg Splits Support Across Four OpenAI Models to Resolve Up to 65% of Routine Inquiries Without a Human

OpenAI published a customer story about Ringg, a voice-agent platform that automates customer support in India: up to 65% of routine inquiries get resolved without a human, by routing work to four different models based on the task. The routing pattern is more interesting than the headline number.

Building with AI··5 min read

I built a system that publishes video automatically: the first real production failure

Traffic OS's first real public canary was one command away from attributing a private-validation video's metrics to a different, public video. Here's what broke, how I caught it, and how I fixed it.

Architecture··9 min read

A silent bug stalled my content pipeline for weeks: here's how I found it

My agent pipeline still looked technically healthy — the crons ran, nothing errored — but its state machine was stuck: nothing ever moved from 'planned' to 'recorded'. Here's what was missing, why it took me so long to see it, and how I fixed it.