Inspiration
Several timers can tell you when individual steps end without telling you when dinner will be ready. An occupied oven or a step that needs longer can disrupt the rest of the meal. Dinner on Time explores how a context-aware cooking companion could help a person recover without losing work already underway.
What it does
Dinner on Time coordinates dependent cooking steps around one serving target. The cook sees what to do next, reports constraints, compares recovery choices and explicitly confirms readiness. The plan distinguishes actual progress from future estimates.
The guided grain-bowl example begins with rice preparation at 6:00 pm. At 6:05 pm its check is overdue, but the step remains running. Reporting the oven unavailable for fifty minutes moves the computed finish from 6:53 pm to 7:30 pm. The original 7:00 pm target and recorded start remain unchanged, while eleven future placements move.
The cook deliberately selects a later target, gives the overdue step more time, confirms it ready and reloads a saved recap. A timer never marks food done. The demonstration confirms one of twelve steps; it does not represent a completed cooked meal.
This interaction suits an Alexa+ concept because it combines a short request, retained context, an understandable consequence and a deliberate next decision. The current working implementation is a web simulation with typed commands and on-screen controls.
How it is built
JavaScript implements a deterministic planner with task dependencies, one cook, equipment capacity and authored attention intervals. The interface uses HTML and CSS. Three menus provide illustrative planning templates.
The standalone app runs locally without an account, model API or runtime packages. Connected mode uses Node.js and six Streamable HTTP MCP tools for meal creation, inspection, recovery comparison, preview, apply and next-step requests. The browser displays actual local tool activity, and the server binds only to loopback. Both modes share the planner.
The project was developed with AI coding assistance. Its command interpreter supports a bounded vocabulary. This submission uses the permitted Alexa+ simulation route; it is not a deployed Echo skill or a live Alexa service.
Challenges and accomplishments
Recovery must preserve recorded starts and confirmed work while changing future placements. Stale replies, interrupted connections and changed saves must not silently replace newer progress. The implementation separates preview from apply, checks returned results and supports deliberate save recovery.
Automated tests cover these boundaries, and the release procedure checks both served app modes, real loopback MCP calls and the actual judge archive against reviewed source. The final accompanying validation receipt identifies the exact tested source. Synthetic scheduling examples remain separate from observed kitchen outcomes.
What is next
The next product evidence should come from an unfamiliar person's first attempt and an observed real meal. Recipe assumptions, voice reliability, measured time savings and physical cooking benefits remain unverified. Further work should follow those observations.
Log in or sign up for Devpost to join the conversation.