Paperboy is building an ambient desktop assistant. At every moment, we understand your intent,
anticipate
your needs, and help—before you ask.
Why now? Local inference is finally fast enough for real-time intent recognition. macOS accessibility and
automation APIs have matured. And unlike capability, which costs billions and advances simultaneously
for everyone, context accumulates with linear storage costs and can't be copied.
Problems we're solving:
- Lightweight, secure capture: Context lives everywhere: screen activity, local
files, apps and APIs. We want to process all of it. But if Paperboy drains battery, hogs CPU, or
leaks data, none of the rest matters. If we get any of these wrong, our product is dead before the
interesting problems start.
- Sub-200ms intent inference: Keystroke at 1:16pm tells us what someone did, not why.
Paperboy needs local models that infer intent from event streams fast enough to be useful—local
inference, careful batching, knowing when to call out to larger models, staying under 200ms for the
interactions that matter.
- Action APIs for agents, not humans: Today's CUI action interfaces are fragile.
Coordinates break when UIs change, element IDs shift between page loads. Agents that open browser
tabs steal your focus. Paperboy has to generalize across UI variations and execute actions in the
background without interrupting you.
- Memory that updates itself: Agents need two kinds of memory: task state that lets
them pick up where they left off, and world knowledge that updates as things change. "Mira was a
prospect; now she's a customer." "Project X was active; now it's abandoned." We're building the
infrastructure that lets agents construct and maintain their own understanding of your world.
- Interpretable context boundaries: Not everything you do is context for everything
else. Your Netflix binges shouldn't inform work recommendations. Your work stress shouldn't set the
tone when you're texting friends. This requires boundary rules that are easy to set, hard to
accidentally cross, and smart enough to handle the gray zones.
- Trust with fine-grained scope: "Help me manage my email" shouldn't mean permanent
access to everything. It means specific senders, time periods, projects. Agents will make
mistakes—the question is whether you can catch them before they're costly, understand what happened,
and roll back.
- Learning from use: Every interaction is training signal. You edit a draft—the
original was wrong. You retry with different phrasing—first attempt failed. You accept
immediately—good output. We're building the feedback loops that extract learning from normal work,
not explicit ratings.
The name "Paperboy" is about delivery—bringing what you need to your door, every day, without you having
to ask. The old paperboy knew your house, your schedule, where to leave it when it rained. That's the
relationship we're building: reliable, personal, and better the longer it lasts.
If these problems are your problems, come build with us! Reach out at john@paperboy.ai.