AI Engineer NYC 2026
Join AI Engineer and Major League Hacking (MLH) on Sunday, October 11, for a full day of building, experimenting, and meeting fellow AI engineers.
Pip
Pip is the Siri that actually gets work done on your Mac, for busy people who live on their laptop.
About the project
Most assistants stop at an answer. Pip keeps going until the job is done.
Pip is a Mac assistant you summon with a hotkey or your voice. It reads the live web in a single call through Nimble, plans and answers with Gemini, and splits bigger jobs across parallel Agent37 cloud computers, then brings the results and files back to your Mac. It also watches pages for you, and instead of storing the pages, it uses Gemini to embed what changed.
Built with
AI tools
Prize Categories Entered
- Submitted
Best Use of Gemini API
Gemini is Pip's brain, called directly through the Gemini API rather than through a third-party model router, and it does seven jobs: 1. Voice: Gemini Flash-Lite turns spoken request into text. 2. Fast path: for small talk, Flash-Lite replies instantly and the planner is skipped. 3. Planning: Gemini Flash turns each request into structured JSON steps (milestones), each marked web read, cloud job, Mac action, or watcher, with dependencies so independent steps run in parallel. 4. Answers: Gemini merges the results into a one-line headline and a comparison table, linking only to sources it actually read. With a screenshot, it also answers "what's on my screen?" from the image. 5. Field extraction: Gemini pulls structured fields like price and status out of the pages Pip watches. 6. Change embeddings: this is the novel part. Gemini Embedding embeds what changed on a page, not the page itself, and alerts are ranked by cosine similarity again
- Event prizeSubmitted
Best Use of Nimble
Nimble is Pip's default way of reading the web; a browser is only the backup. Every read step goes to Nimble's Search and Extract APIs, fanned out in parallel: "compare this product at five stores" is five simultaneous Nimble calls instead of one browser clicking through five sites. Only if Nimble comes back empty, blocked or thin does the router escalate to a slower cloud browser (an Agent37 worker), and every answer records which one produced it. Nimble also powers the watchers: Pip snapshots pages with Nimble Extract on a timer, and those snapshots are the "changes" Gemini embeds. To back the speed claim with evidence, Pip includes a benchmark that runs the same five tasks on Nimble and on a cloud browser and reports median time and success rate, so the speedup is measured, not asserted.