MentorFlow

Deployable Link: https://mentorflow-app-dhonj.ondigitalocean.app

Technology can deliver information. People create understanding.

A calm teaching companion for tutors, mentors, volunteers, and older siblings — the people who teach one learner at a time. MentorFlow is a private teaching journal with intelligent structure: you reflect for two minutes after a session, and DigitalOcean Gradient AI turns that reflection into durable memory — how this learner learns, what to try next, and a warm summary for their family.

The loop

Prepare → Teach → Observe → Adapt → Reflect → Repeat

How Gradient AI is used (three surfaces, one key)

Feature	Gradient surface	Model
Reflection engine — turns a raw post-session reflection into learning patterns, next-session moves, and a family summary	/v1/chat/completions	openai-gpt-5.5
Lesson planner — drafts a session plan grounded in the learner's accumulated memory	/v1/chat/completions	openai-gpt-5.5
Family translation — parent summary in the family's home language	/v1/chat/completions	openai-gpt-5.4-mini (right-sized: cheap model for light work)
Voice notes — the family summary as audio, in their language	/v1/async-invoke (submit → poll → fetch)	fal-ai/elevenlabs/tts/multilingual-v2

Built on DigitalOcean Gradient AI Serverless Inference — one Model Access Key, the OpenAI-compatible Chat Completions API, and structured JSON output (every endpoint prompts for and parses a strict JSON shape, e.g. the debrief returns {observations, timelineEntry, parentSummary, nextFocus, recommendations, homework, strategyNote, trendNote}).

Model flexibility is a feature, not an afterthought: the deep-synthesis model, the fast translation model, and the TTS model are each a single env var, so you can swap any model from the Gradient catalog without touching code. We default to one strong model (openai-gpt-5.5) for judgment-heavy synthesis and a smaller one for light language work.

The debrief loop is the core: mentor marks a session done → fills a quick structured debrief (what clicked, what confused, strategy worked/didn't, confidence + engagement ratings, review-next, and an opt-in parent summary) → Gradient AI returns the teaching intelligence above → the learner's memory, trends, timeline, and recommendations update. The parent summary intentionally does not exist until after a debrief — before the lesson, the app doesn't know what happened.

No API key? Every AI feature falls back to believable mock output — the demo cannot die on stage.

Run it

npm install
cp .env.example .env   # add your DigitalOcean model access key
npm run dev            # app on :5173, API on :3001

.env:

DIGITAL_OCEAN_MODEL_ACCESS_KEY=your-key
DIGITAL_OCEAN_MODEL=openai-gpt-5.5        # optional
DIGITAL_OCEAN_MODEL_FAST=openai-gpt-5.4-mini  # optional

Deploy (DigitalOcean App Platform)

1. Push to GitHub.
2. App Platform → create app from the repo (Node buildpack).
3. Build command npm run build, run command npm start.
4. Set env vars DIGITAL_OCEAN_MODEL_ACCESS_KEY and PORT=8080.

Express serves the built frontend from dist/ in production — one component, no separate static site.

Demo (90 seconds)

1. Dashboard — today's sessions, learners, what MentorFlow is noticing. One primary action: Prepare next session.
2. Prepare Jasmine's fractions session — the plan is drafted around what MentorFlow remembers about her (pizza model works, number line doesn't yet).
3. Teach (offstage). Then reflect: two honest minutes about where she lit up and where she struggled.
4. Watch her profile rewrite itself — patterns, next-session moves, family summary.
5. One click: the summary in Spanish. One more: as a voice note her family can listen to.

Stack

React + Vite · Express · DigitalOcean Gradient AI · a JSON file as the database (18-hour hackathon; the database is not the point)

Reset demo data

curl -X POST localhost:3001/api/reset