What we learn running AI on our own GPUs

Opus Plans, Sonnet Executes, Fable Advises: What We Learned Running a Team of Claude Code Agents

A post on X laid out a cheaper way to run Claude Code: the strong model plans, a mid-tier model executes, and an advisor speaks only when needed. This is the setup for one session, what…

14 min read

Ten AI Agents, One Studio: How We Build Optra Studio and the Rules That Keep the Agents Honest

Optra Studio makes an animated series with no crew, and ten Claude agents build and run it at once. What each agent owns, the rules that stop them shipping work that only looks finished…

9 stills

What Is Inference Latency? Definition, Causes, and How to Measure It.

6 min read

How a Shot Is Made: From One Line of Script to a Checked Picture on Screen

Every shot in an Optra Studio film goes through the same road: words, checks, a picture, more checks, a moving clip, every frame read, and a place in the cut. This is that road, step by…

13 min read

NPU vs GPU: How They Compare for AI Computing

5 min read

AI Render: The Complete Guide to Rendering Using AI

4 min read

How to Choose a GPU for AI Inference in 2026

6 min read

The Edit Runs Itself: Cutting an Animated Series With Nobody in the Room

We are making a fifteen-episode animated series, in Arabic and English, and no person cuts it. There is no editor, no timeline anyone scrubs, and no approval step between the last shot…

6 min read

Tuning a 27B on Two RTX A5000s: What We Measured, What Worked, and What Did Nothing

A measured study of a self-hosted Qwen3.8-27B on two RTX A5000s under vLLM: the real throughput curve, the three different limits on concurrency, a free 2x from prompt ordering, and two…

11 min read

Casting Seven Voices Without Casting Anyone: Designed References for an Animated Series

We are making a fifteen-episode animated series set in a UAE village, and it has eleven speaking parts: two children, their parents, a grandmother and grandfather, and the neighbours along…

50 voice takes

Writing a 15-Episode Animated Series on Our Own GPUs: A 27B Model, Two Passes and a Calculator

We are making a fifteen-episode animated series, about twenty minutes an episode, entirely on hardware we own. This post is about the least glamorous part of that and the part that turned…

3 min read

How To Speed Up LLM Inference: A Practical Guide

If you've deployed a language model in production, you already know the pain: responses that lag, GPU bills that climb, and users who notice every extra second of latency. LLM inference is…

5 min read