The build journal

How we build Scroll, decision by decision.

Not the marketing arc — the decision arc. Field notes on what we bet on, what we got wrong, and what shipping taught us. Newest first; read bottom-up for the whole story so far.

  1. Costs10 min read

    Managing API costs in Scroll

    Scroll is a one-time Mac app, but cloud speech uses your own OpenAI API key. Here is what each mode sends, what models it uses, and how to keep the bill boring.

    A practical guide to Scroll's three modes: Look, Ask, and Read Aloud. Learn where to put your OpenAI API key, which models each mode uses, and how to choose the cheapest path for each reading job.

    Read the post →
  2. Architecture12 min read

    The next decision: splitting intelligence from voice

    Why we're moving regular narration off the realtime model — for a 7–15× cost cut, better explanations, and a simpler pipeline the architecture was already shaped to allow.

    Realtime speech-to-speech is brilliant for conversation and the most expensive thing we buy. Narration isn't a conversation — so we're decomposing it into a cheap text LLM plus a cheap streaming TTS, and keeping realtime only where it earns its price.

    Read the post →
  3. Design11 min read

    Highlighting what the voice is saying — and the dead ends on the way

    Read-along highlighting looks like a UI flourish. It's really a matching-and-timing problem, and the version that works does less than the versions that didn't.

    We wanted the page to light up in step with the voice. One approach fought the audio API and lost. The methodology that stuck: deterministic transcript matching, audio-paced reveal, and a model-called layer that's optional, not load-bearing.

    Read the post →
  4. Product11 min read

    We decided to stop interrupting you

    The plan said cancel narration on every scroll. Shipping it proved the plan wrong — and the same review hardened how Scroll handles memory, privacy, and trust.

    Driving stale audio to zero meant interrupting the voice mid-sentence on every scroll. It felt worse, not better. Here's the reversal — plus the methodology behind document memory and the 'private by architecture' stance it shipped alongside.

    Read the post →
  5. Engineering13 min read

    Locking a voice to your eyes

    How Scroll reads the screen, figures out where you're looking, decides when to talk, and stops cleanly the instant your eyes move on — the methodology behind the core loop.

    A guided tour of Scroll's core loop: confidence-scored screen capture, a model of your attention, dwell-and-settle pacing, a raw WebSocket to the model, and an audio ledger that makes interruption honest. Methodology, not tuning constants.

    Read the post →
  6. Origin8 min read

    Reading became the job. So we built a way to listen to it.

    Why Scroll exists, the one idea everything else is built on — your scroll gesture is the input — and the constraints we accepted before writing a line of narration code.

    Scroll didn't start with a feature. It started with a problem we couldn't out-work — too much important reading, and no good middle ground between text-to-speech and summaries. This is the founding bet, and the rules we set before building.

    Read the post →

← Back to Scroll