WARM

Inspiration

Households with solar panels, a home battery and an electric car can see more energy data than ever, yet still face one unanswered question: what should I do next?

Most energy dashboards report the past. Device apps optimize one component at a time. But a home behaves as one connected system: car charging affects the battery reserve, solar generation changes the best timing, tariffs affect grid use, and comfort or departure needs can invalidate a financially attractive schedule.

I work in supplier and contract management rather than energy or software engineering. WARM grew through repeated ChatGPT and Codex conversations in which I challenged household documents, energy assumptions, tariffs, charging evidence, privacy boundaries and product choices. OpenAI Build Week transformed that accumulated learning into a separate, public and privacy-safe competition artifact.

What it does

WARM turns synthetic solar, household demand, battery reserve, electricity-price and EV-charging signals into understandable next decisions.

The main product journey contains four connected views:

Today

Today recommends two useful times and exposes the complete reasoning chain:

signal → rule → recommendation

It also shows the modeled daily value and the conditions that would change the advice.

Insight

Insight reconciles one fictional 365-day household. It separates:

• synthetic meter-like values;
• inferred calculations;
• explicit assumptions; and
• forecasts.

Storage movement is kept separate so energy is not counted twice.

Decide

Decide ranks understandable next steps by:

• modeled opportunity;
• confidence;
• household effort; and
• evidence that could change the ranking.

Potential savings that may overlap are explicitly not added together.

Explore

Explore lets the user change generation, flexible demand and storage assumptions. Every result remains clearly labelled as hypothetical.

The repository also contains a condensed 90-second interactive jury experience. A judge can change the car departure time, minimum battery reserve and tomorrow’s scenario, after which WARM immediately recalculates the recommendation and expected outcome.

Both experiences advise and explain. They never send commands to a charger, battery, inverter or other device.

How we built it

The public artifact uses:

• Node.js 20;
• dependency-free HTTP servers;
• bundled synthetic JSON;
• browser-native HTML, CSS and JavaScript; and
• deterministic tests.

It requires:

• no package installation;
• no account or API key;
• no database;
• no third-party SDK;
• no external API;
• no persistence; and
• no equipment-control integration.

The tests launch both submitted experiences, load the required routes and assets, reconcile the fictional energy identities, exercise the scenario calculations, confirm the synthetic-data and no-device-control disclosures, and reject unauthorized mutation requests.

The video-aligned experience runs with:

npm test
npm run start:video

Then open:

http://127.0.0.1:4175/vandaag.html

How Codex and GPT-5.6 helped

I retained the consequential product decisions: comfort before savings, advice rather than actuation, visible uncertainty, privacy-safe publication and recommendations that remain understandable and reversible.

Codex with GPT-5.6 accelerated the work by:

* translating the connected-home concept into the Today → Insight → Decide → Explore journey;
* implementing the coordinated recommendation and interactive scenario controls;
* challenging energy double counting and false precision;
* keeping modeled outcomes separate from measured facts;
* designing visible reversal conditions and rejected alternatives;
* building the video-aligned demo and condensed jury narrative;
* creating privacy, network, persistence and mutation boundaries; and
* preparing reproducible setup, tests and submission evidence.

The collaboration was iterative. I repeatedly challenged results that did not make household sense or crossed the agreed risk boundary. Codex supplied implementation speed, adversarial critique and deterministic proof.

Challenges

The hardest problem was preventing a plausible-looking number from becoming a false claim.

Energy flows must reconcile. Storage movement must not be double-counted. Alternative savings can overlap. Forecasts must look different from facts. A scenario must not masquerade as a prediction, tariff quote or equipment command.

The public repository also had to demonstrate meaningful Build Week work without exposing private household records or supplier-specific integrations.

Each issue became a visible qualifier, an explicit model boundary or a deterministic test.

Accomplishments

* A working Today → Insight → Decide → Explore product journey.
* A visible signal → rule → recommendation trace.
* Explicit reversal conditions for every important recommendation.
* One internally balanced fictional 365-day household.
* A bounded interactive what-if calculation.
* A second interactive 90-second jury experience.
* A dependency-free and credential-free architecture.
* No persistence or equipment-control path.
* Deterministic validation covering both submitted experiences.
* A public MIT-licensed repository aligned with the public video.

What we learned

For household AI, uncertainty design is product design.

Advice becomes more useful when the interface explains what was supplied, derived, assumed and forecast — and what evidence would change the decision.

AI can also let a determined non-specialist cross disciplinary boundaries, but only when curiosity is paired with provenance, tests, exact versions and a willingness to reject plausible-looking answers.

What’s next

The next step is an opt-in verification loop: the household accepts one recommendation, observes the result after enough time and compares it with the original estimate.

Any future use of private data or equipment integration would remain optional, consent-based and outside this public competition artifact.

Built With

• codex
• css3
• generative-ai
• github
• gpt-5.6
• html5
• javascript
• node.js