Live demo: https://stepwise-8ev4.onrender.com (free tier — first load after idle takes ~1 minute) · demo login [email protected] / demo-password-123
Source code: https://github.com/favour187/nextstep-hacks-2026
What it is
StepWise is an AI sustainability coach that turns a vague environmental worry ("I throw away too much plastic") into a concrete, measurable action plan — and then tracks it.
The problem it attacks: most climate-concerned people never act because broad goals don't produce behaviour. StepWise reframes the worry into a target the person controls, scores candidate actions with a real decision model (benefit vs. friction: effort, cost, time, habit, learning), works out what is actually feasible given the person's time and budget, and then coaches them day to day — with cumulative impact (waste kg, CO₂e, kWh, water litres, food kg, items) and a consistency streak.
Why it's interesting (and honest about it): the recommendation logic is a deterministic, unit-tested decision model — not the LLM's opinion. The model is inspired by the behavioural-science concepts of sludge (friction that blocks intended action) and the intention–action gap. The LLM layer makes the same numbers conversational; with no API key configured the app runs fully on deterministic local skills.
Features
Reframe — your concern becomes one specific behaviour target with a clearly named lever. Decision-scored plan — 28 curated actions across 6 categories (waste, energy, water, transport, food, consumption), each with quantified impact and friction attributes; scored and ranked deterministically. Feasibility analysis — how many times each action can realistically be done within your committed time and budget over the chosen horizon. Milestones — a schedule from "learn one fact" to the 30-day review. Daily check-ins — one-tap logging; impact meters update cumulatively, with a 7-day activity chart and streak. AI coach — chat that explains why an action was chosen (it cites the scoring model), unblocks you when stuck, and reviews progress. Falls back to deterministic local skills when no API key is set. Auth & persistence — email/password accounts (PBKDF2), opaque session tokens, per-user plans.
Honest notes (for judges & reviewers)
Impact estimates are order-of-magnitude, typical-value figures chosen for teaching, not precision claims. The README + UI label them as estimates; the architecture makes it trivial to swap the library.py numbers for region-specific data (e.g. EPA/UK DEFRA factors).
No greenwashing: StepWise measures your actions, in units you can actually verify (kWh on a bill, litres per minute, kg per bag), rather than projecting world-saving abstraction AI disclosure: AI was used for debugging using Claude sonnet which was used under the my direction.
Built With
- claude-sonnet
- css
- flask
- html
- javascript
- python