Task Hero

Speak your campus need, AI structures it in 10 seconds, a hero grabs it and gets paid. Real-time task marketplace powered by GPT-4o-mini and Deepgram, built in Flutter, live and deployed at SUTD.

Comment

💡 Inspiration

Every day at SUTD, the same scene plays out — a student stuck in a lab at 2am needs food but can't leave, someone's parcel is sitting at Building 1 reception but they're in class across campus, a freshman is drowning in Python homework and just needs 30 minutes of help.

The campus is small. The community is tight. But there's no easy way to say: "Hey, I need a hand — and I'll pay you for it."

WhatsApp groups get chaotic. Telegram chats scroll past. We wanted something purpose-built — a marketplace where asking for help takes 10 seconds, not 10 messages.

🚀 What it does

TaskHero is a real-time campus task marketplace where SUTD students post tasks and other students (Heroes) complete them for micro-compensation.

🎤 Speak or type your task → 🤖 AI instantly formats it with title, category, location, and fair pricing → 📋 It goes live for Heroes to grab → 🏃 Real-time tracking shows both parties live progress → 💰 Payment released on completion

The flow:

Step What Happens
1️⃣ Poster describes a task (voice or text)
2️⃣ GPT-4o-mini structures it — title, category, location, suggested compensation
3️⃣ Task goes live on the marketplace
4️⃣ A Hero accepts → both see real-time progress updates
5️⃣ Hero completes → Poster confirms → Payment released (95% to Hero)

Key features:

  • 🎙️ Voice-to-Task — Speak naturally, Deepgram transcribes, AI formats
  • 📊 Live Dashboard — Real-time stats, active missions, activity charts
  • 🔍 Smart Browse — Filter by category, sort by price/urgency, keyword search
  • 📱 Fully Responsive — Mobile bottom-nav ↔ Desktop sidebar, works on any screen
  • ⚡ Instant Updates — Firestore streams power every screen in real-time

Live demo: taskhero-sutd.web.app

🛠️ How we built it

Flutter Web (Dart)  →  UI & responsive layout
ShadCN UI           →  Polished component library
Firebase Auth       →  One-click Google Sign-In
Cloud Firestore     →  Real-time database + streams
OpenAI GPT-4o-mini  →  Task formatting AI
Deepgram Nova-2     →  Speech-to-text engine
MediaRecorder API   →  Browser mic capture (WebM/Opus)
Firebase Hosting    →  Deployed & live

Architecture in 30 words: User speaks/types → Deepgram transcribes → GPT-4o-mini formats into structured JSON → Firestore stores it → Streams push updates to all connected clients in real-time. Zero polling.

We used Firestore transactions for task acceptance (no double-booking), FieldValue.increment for atomic counter updates, and 7 real-time streams powering the dashboard, browse screen, task detail, and profile.

Built in under 24 hours. ~3,500 lines of Dart across 14 files.

🧗 Challenges we ran into

  • Google Sign-In + Cross-Origin headers — Firebase Auth popups break if Cross-Origin-Embedder-Policy is set to require-corp. Took us 2 hours to discover credentialless was the fix. One word. Two hours.

  • Firestore on Flutter Web — The default gRPC transport doesn't work on web. Had to switch to long-polling, which isn't documented prominently.

  • Voice recording across browsers — MediaRecorder API support varies wildly. We had to handle codec negotiation (WebM/Opus vs fallbacks) and deal with empty audio buffers on some browsers.

  • AI response consistency — GPT-4o-mini occasionally returned creative field names instead of our exact schema. Solved with response_format: json_object and strict schema instructions.

  • Double-posting bug — Users tapping "Post Task" fast enough would create duplicate Firestore documents. Fixed with an isPosting mutex guard and button disable state.

🏆 Accomplishments that we're proud of

  • ⚡ < 3 second task creation — from voice to posted task on Firestore
  • 🎤 Voice-first UX — most campus apps don't even have text input done right, we shipped speech-to-text
  • 🔄 True real-time — both poster and hero see the same live-updating task detail screen, zero refresh needed
  • 🤖 AI that actually helps — not a gimmick; GPT-4o-mini genuinely saves users from writing structured posts, suggests fair SUTD-context pricing ($3–4 for canteen runs, $10–20 for tutoring)
  • 📱 Responsive from day one — not "it works on mobile", but proper mobile-first bottom nav ↔ desktop sidebar transitions
  • 🔒 Production-grade security — Firestore rules enforce user-level access, transactions prevent race conditions, API keys properly gitignored

📚 What we learned

  • Firebase on web is a different beast — half the Flutter+Firebase docs assume mobile. Web has its own quirks (long-polling, CORS headers, popup auth vs redirect auth).
  • AI as a formatting layer, not a feature — The best AI integrations are invisible. Users don't care that GPT-4o-mini is structuring their task — they care that it takes 10 seconds instead of 2 minutes.
  • Real-time is addictive once you have it — Once we wired up Firestore streams, we couldn't go back to fetch-and-refresh. Every screen became a live dashboard.
  • Voice UX requires patience — Recording, transcribing, formatting is a 3-step async pipeline. Each step can fail. Error handling isn't optional — it's the product.

🔮 What's next for TaskHero

  • 📲 Push notifications — Alert heroes when a new task matches their preferences
  • 💬 In-app chat — Let poster and hero coordinate without leaving the app
  • ⭐ Reputation system — Build trust scores based on completion rate, ratings, response time
  • 🏫 Multi-campus expansion — NUS, NTU, SMU — same compact campus model, much larger user base
  • 💳 Real payments — Integrate PayNow/Stripe for actual SGD transfers instead of honor-system tracking
  • 🧠 Smart matching — AI recommends tasks to heroes based on location, skills, and availability

Built with 🧡 by SUTD students, for SUTD students.

🔗 Live App · 💻 GitHub

Built With

  • dart
  • deepgram-speechtotext
  • firebase
  • firestore
  • flutter
  • hosting
  • openai-gpt4o-mini
  • shadcn-ui
Share this project:

Updates