Inspiration
An evening changes faster than a timetable. Dinner moves, laundry can wait, and a bike still needs charging. Watt If explores a focused question: can a conversational assistant help rearrange optional chores while keeping the person in control?
What it does
Watt If plans three explicitly deferrable chores in a fictional household. A person describes a change, inspects the interpreted start times, deadlines and quiet hours, then reviews a checked timetable. An infeasible request remains blocked. A correction produces a new revision that needs its own approval. Approval unlocks a real downloadable advisory calendar containing only that approved schedule.
The demonstration includes a successful deadline request, a conflict between an exact dryer start and quiet hours, and a correction that preserves the exact start while releasing the old deadline. Both calendar downloads are actual artifacts. No calendar is imported and no appliance is operated.
How it is built
The Python application uses Strands Agents SDK with Qwen 2.5 14B through Ollama. Bounded tools translate source-grounded instructions into constraints and call the existing PeakShift exact scheduler. The model cannot approve a proposal or export a calendar. The server enforces revision, request and content-hash checks.
The planner evaluates the finite half-hour scheduling space and compares proposed timing against the same included chores at their earliest allowed starts. Peak, tariff and carbon figures are fictional modeled comparisons. Skipped work is explicitly omitted rather than counted as an efficiency improvement.
Pending requests survive failed interpretation and restarts. Follow-up requests retain unresolved obligations, while explicit manual replacement provides a recovery route. Exact starts are separate from not-before windows and finish deadlines.
The frontend uses plain HTML, CSS and JavaScript. A textured midnight surface, credited marble art and original clock geometry lead into a cobalt request panel and paper timetable. Typography is self-hosted. The schedule, conflict, approval and download states belong to the working application.
Verification
The current suite passes 58 regression tests, including SDK fixtures, source-grounding guards, request recovery, stale approvals, exact-start behavior and HTTP boundaries. An independent scheduler check covers 3,546 scenarios and 140,042 assignments with zero oracle mismatches; 29 workflow checks also pass. These are finite checks, not a general-language reliability percentage.
Fresh demonstration evidence uses the same Qwen model digest through an explicitly configured RunPod SSH loopback forward. The interface and trace disclose that runtime. The app has no automatic fallback, cloud-account credential or GPU-rental function. Local model startup failures and unsupported-command failures remain documented separately from successful warmed execution.
Reused work and limitations
PeakShift is pre-existing work by the entrant. Watt If substantially extends it with conversation, persistence, recovery, checked constraints, a working visual interface, approval and calendar export. Source provenance and changes are documented in the repository. The solver is unchanged.
This is the competition's simulated Alexa+ route. It has no Alexa connection, voice integration, live utility data or appliance control. All household data is fictional; calendar events use October 1, 2026 as floating local times. The application supports a deliberately small command vocabulary, so unsupported wording can require clarification or manual controls. Inspected constraints remain essential.
Demo and credits
The screen recording shows actual execution at normal speed. Synthetic ElevenLabs Jessica (eleven_v4) narration is presentation, not an app feature. A brief, visibly labeled Higgsfield illustration introduces the film. Thorvaldsen's Night photograph is credited under its recorded CC0 license; fonts retain their SIL Open Font Licenses. AI assistance and asset provenance are disclosed in the repository.
Next work would evaluate more independently selected language cases and improve clarification around unsupported requests. Device integration would need separate permissions and real integration testing.
Source and demonstration
Source, setup and actual execution artifacts: https://github.com/Bembaby/watt-if
October 21 demonstration: https://youtu.be/Oq7R-jTPYzM
Recorded executing source and approved upgrade commit: d242ec3aaa6f1ab5d195a5facc68cf49f8afcb5c.
Log in or sign up for Devpost to join the conversation.