Mise

Inspiration

AI workflows involve repeated waiting between steps. We wanted to use that time to prepare what the user would likely need next. Mise comes from mise en place: “everything in its place.” Chefs prepare ingredients before cooking; Mise prepares likely results before you ask—a natural fit for the hackathon’s cooking theme.

What it does

Mise predicts your next action and prepares its result in the background. After reading a receipt, it might prepare field extraction while you review the text. A Ready badge appears before you click. Mise also adapts to repeated choices, such as learning that you prefer summarizing whiteboards over creating tasks. The dashboard shows predictions, skipped calls, learning changes, and actual waiting times. Note: Mise runs locally. The GitHub README includes setup instructions; live AI features require Gemini and Perceptron API keys. A working live demo will be available during judging.

How we built it

We used React and Vite for the frontend and Node.js with Express for the backend. Perceptron reads images, while Gemini extracts fields, summarizes text, and creates tasks. A lightweight Markov model predicts the next action using generated training journeys and choices from the current session. Confidence thresholds, estimated costs, and request budgets control prefetching. Caching and request sharing prevent duplicate calls.

Challenges we ran into

Incorrect predictions can waste paid calls, so we had to balance responsiveness with cost. We also handled early clicks, expired results, and provider failures while keeping readiness indicators accurate. Fair benchmarking required fresh sessions, identical actions, and equal review time.

Accomplishments that we're proud of

Connected real AI calls to working prediction and caching. Made background preparation visible before the next click. Demonstrated predictions adapting to user choices. Recorded 3.46 s → 2.60 s median total action waiting across eight matched, scripted US-receipt benchmark groups.

What we learned

Responsiveness depends on when work starts, as well as how quickly it finishes. We learned that skipping an unlikely call can be as valuable as preparing the correct one—and that unused calls and failures belong alongside speed improvements in our measurements.

What's next for Mise

We want to build a small integration SDK, evaluate consented real-world workflows, improve cost estimates, and add optional persistent personalization. Our goal is to bring mise en place to more applications: the next result, prepared when you need it.