StepWise

• 

• 

• 

  App does not give any direct answers to problems.

---

🧠 Inspiration

Every student knows the feeling: you copy a solution from Google, feel fine about it, and then completely blank on the test.

The problem isn't access to answers — it's that answers without struggle produce zero learning. I built StepWise because I wanted to fix that. Not with another flashcard app or quiz tool, but with something that actually forces you to think.

🔍 What it does

StepWise is an AI-powered homework tutor that guides students through any problem one step at a time — and never gives the answer, no matter how they ask.

When a student submits a problem, StepWise doesn't solve it. Instead it identifies the problem type, orients the student, and gives exactly one concrete first step. When the student responds — right or wrong — it acknowledges their attempt and gives the next single step. The conversation continues until the student has worked through the entire problem themselves.

🚫 If they try to ask for the full solution, a guardrail fires and redirects them: "I won't give you the answer — but here's your next step."

The tutoring is subject-aware with four distinct styles:

Subject	Tutoring style
Math	Identifies problem type and entry method before any computation
History	Pure Socratic method — only asks questions, never states facts
English	Anchors to thesis, structure, and argument
Science	Grounds students in knowns vs unknowns before touching any formula

After finishing a problem, students see a session recap: the concept they practiced, a personalized tip for next time, and a 1–5 confidence rating. Everything saves to a progress dashboard tracking their streak, average confidence, subjects covered, and hardest concept over time.

Teachers get a fully separate password-protected dashboard showing every student's sessions in real time — concepts struggled with, steps taken, average confidence, and a cheating-risk score. A class heatmap shows at a glance which concepts are easy or hard across the whole class. Teachers can create assignments, manage their roster, and generate progress reports in one click.

⚙️ How I built it

Layer	Tech
Frontend	React 18 + Vite, dark glass morphism design system in CSS-in-JS
AI	OpenAI Responses API with five subject-specific system prompts
Backend	Vercel serverless function — API key never touches the client
Auth + Storage	Supabase for signed-in users, localStorage for guests
Deployment	Vercel — live public URL, no setup required

🚧 Challenges I ran into

Getting the AI to give exactly one step — not two, not a veiled answer — took significant prompt iteration. Early versions would give "step 1" and "step 2" in the same message, or sneak the answer in by being "too helpful."

The harder problem was detecting answer-seeking without over-blocking. Students try creative phrasing to extract answers — distinguishing that from a legitimate question, without the guardrail firing on innocent messages, required careful engineering.

On the teacher side, the challenge was making the dashboard genuinely useful rather than decorative. Surfacing signals a teacher could actually act on — not just session counts, but concept-level struggle patterns and cheating-risk flags — required thinking carefully about what information actually changes teacher behavior.

🏆 Accomplishments I'm proud of

The guardrail system works reliably and gracefully — it never shames students, it just redirects them.

The subject-specific tutoring styles produce noticeably different, more appropriate responses depending on what the student is working on.

And the teacher dashboard surfaces real, actionable intelligence:

Not just "Miles did 3 problems" — but "Miles struggles with WWI origins, took an average of 7 steps per problem, and has a Medium cheating-risk score."

That's something a teacher can actually use on Monday morning.

📚 What I learned

Prompt engineering is the core product in an AI tutoring app — more so than the UI or the infrastructure. A single sentence change in a system prompt can mean the difference between a tutor that genuinely teaches and one that just paraphrases the textbook.

I also learned that in EdTech, the teacher view is often what closes the sale. Showing that a tool fits into an existing classroom workflow — not just a student's solo study session — is what makes schools willing to pay for it.

🔮 What's next for StepWise

• Assignment mode — teachers push a specific problem to the whole class and monitor live as students work through it
• Join codes — real classroom accounts where teachers invite students directly
• Parent dashboard — weekly progress emails summarizing what their child worked on and where they struggled
• Mobile optimization — full responsive layout for students on their phones
• LMS integrations — Google Classroom and Canvas so StepWise fits into workflows schools already use

Built With

• css
• html
• javascript
• open-ai-api
• react-18
• supabase
• vercel