Skip to content

Latest commit

 

History

History
280 lines (213 loc) · 16.9 KB

File metadata and controls

280 lines (213 loc) · 16.9 KB

Hackathon Idea Research — 24-Hour Build

Team: All Full-Stack Developers

Themes: Education | Healthcare | Re-engineering


CHOSEN IDEA: VisualDebugger — AI Bug Tutor for VS Code

"Vibe coding makes you fast. VisualDebugger makes you fast AND smart."

The Concept

A VS Code extension for students and developers that turns every AI-assisted bug fix into a learning moment. When a user asks AI to fix code, VisualDebugger doesn't just apply the fix — it shows a live visual preview, classifies the bug, explains why it broke, and tracks your personal bug patterns over time.

The Problem

  • "Vibe coding" is the hottest and most controversial trend in dev education — students ship code they don't understand
  • AI tools fix bugs silently, creating a dependency loop where developers never learn from their mistakes
  • Developers spend ~70% of debugging time reproducing and understanding bugs, not fixing them
  • Students accept AI-generated fixes without building mental models of what went wrong

The Narrative (Pitch Angle)

AI coding tools made students faster but not smarter. Every accepted AI fix is a missed learning opportunity. VisualDebugger is the missing feedback loop — every bug you hit becomes a lesson you retain. It's not anti-AI; it's AI that teaches instead of just doing.

Theme fit: Education — AI-powered learning tool that turns every bug into a teachable moment.


Core Features (The Full Loop)

Feature What It Does Why It Matters
AI Fix + Visual Diff Shows exactly what changed in the code with highlighted diff Baseline — user sees the change, not just the result
Live Preview Split panel: broken UI on left → fixed UI on right (for web projects) Makes the fix concrete and visual, not just abstract code
Bug Classification Labels the bug type: "CSS Specificity", "Off-by-one", "Null Reference", "Async Race Condition" Gives students vocabulary for bugs — naming things is half of understanding
"Why It Broke" Explanation Plain-English root cause + the underlying concept e.g., "Array.map returns a new array — you treated it as mutating in place. This is a common mistake with functional array methods."
"Test Yourself" Quiz Before revealing the fix, asks "What do you think was wrong?" Active recall — the most effective learning technique. Turns passive fix-acceptance into engagement
Concept Link Connects the bug to a broader concept with a mini-lesson "Want to learn more? This relates to: Immutability in JavaScript"
Bug Dashboard Tracks personal bug patterns over time — heatmap of your weak spots "You've made 7 CSS layout bugs, 3 async errors, 2 type mismatches this week. Focus area: async/await."

The Magic Moment (Demo)

  1. Open VS Code with a broken React app (button doesn't work, layout is off)
  2. Run the app — show the broken UI in the live preview panel
  3. Ask AI to "fix this" — VisualDebugger intercepts
  4. Instead of silently applying the fix:
    • A "Test Yourself" prompt appears: "What do you think is wrong with this code?"
    • Student guesses (or skips)
    • The fix is revealed with a visual diff + the live preview updates in real-time showing the UI go from broken → fixed
    • A bug classification card slides in: "Bug Type: CSS Specificity — your class .btn was being overridden by a more specific selector .container .btn"
    • A "Why It Broke" explanation appears with the concept
  5. Switch to the Bug Dashboard tab — show a week of simulated coding data:
    • Heatmap: "Your top bug categories: Layout (7), State Management (4), Async (3)"
    • Trend line showing bugs decreasing over time (you're learning!)
    • "Recommended focus: Async patterns in JavaScript"

The "wow": The bug dashboard. Judges see a student's coding weaknesses visualized — it transforms a code fixer into a personal coding coach.


Tech Stack

Layer Technology Why
Extension VS Code Extension API (TypeScript) Native integration, webview panels for UI
Bundler esbuild Industry standard for VS Code extensions, 100x faster than webpack
Live Preview VS Code Webview + embedded iframe (DigitalOcean or localhost) Shows before/after of web app in real-time
Diff Visualization diff (compute) + diff2html (render) Syntax-highlighted side-by-side diffs out of the box
LLM Pipeline Gemini API Bug classification, explanations, quiz gen. Targets MLH Gemini prize
TTS ElevenLabs API "Read aloud" for explanations. Targets MLH ElevenLabs prize
Bug Dashboard Chart.js v4 in VS Code webview panel Canvas-based, no framework dependency, small bundle
Webview UI @vscode-elements/elements (Lit-based) Replaced deprecated webview-ui-toolkit. Native VS Code look-and-feel
State/Storage MongoDB Atlas (primary) + VS Code globalState (fallback) Cross-device persistence. Targets MLH MongoDB prize
Demo Hosting DigitalOcean App Platform Production deployment. Targets MLH DigitalOcean prize
Scaffolding Yeoman generator (yo code) Fast VS Code extension bootstrap

MVP Scope (24 Hours)

BUILD (must have for demo)

  • VS Code extension that detects when AI applies a code fix (intercept diff/file change)
  • Visual diff panel showing what changed with syntax highlighting
  • Live preview panel showing before/after UI state (scoped to web dev — React/HTML/CSS)
  • LLM-powered bug classification (5–8 common categories)
  • "Why It Broke" explanation with plain-English root cause
  • "Test Yourself" quiz prompt before revealing the fix
  • Bug Dashboard with heatmap/chart showing patterns from session data
  • Pre-seeded demo data showing a realistic week of a student's coding bugs

SKIP (not needed for demo)

  • Support for non-web projects (Python scripts, backend-only code)
  • Real persistent storage across sessions (use in-memory + seed data for demo)
  • Integration with specific AI tools (Copilot, Cursor, etc.) — simulate the AI fix trigger
  • Concept library / mini-lessons (just show the concept name + one-liner)
  • Multi-language support
  • Settings/configuration UI
  • Marketplace publishing

Risks and Mitigations

Risk Severity Mitigation
VS Code extension dev is fiddly and slow High Use yo code generator. One person owns extension scaffold full-time for first 4 hours. Test webview early.
LLM explanations are generic/unhelpful Medium Few-shot prompt with 5–6 hand-crafted examples of great bug explanations. Test and iterate on prompts in hours 2–6.
Live preview is hard to get working reliably Medium Scope to web dev only. Use a simple embedded iframe pointing to localhost dev server. Have a fallback: screenshot before/after if iframe is flaky.
"Test Yourself" quiz feels gimmicky Low Make it optional (skip button). Frame it as "active learning mode" in the pitch. Research backs it up (active recall).
Demo breaks live on stage Medium Pre-plant 3 specific bugs in the demo app that you've tested the full pipeline with. Never go off-script. Pre-record backup video.
Bug dashboard looks empty with one session Low Pre-seed with a week of realistic simulated data. This is standard hackathon practice.

24-Hour Build Timeline

Phase Hours Activity Owner Suggestion
0–2 2h Extension scaffold (yo code), project architecture, API keys, divide work All together
2–6 4h Track A: Extension core — file change detection, webview panels, diff rendering Dev 1 + Dev 2
2–6 4h Track B: LLM pipeline — bug classification, explanation gen, quiz gen (test in isolation with sample diffs) Dev 3 + Dev 4
6–10 4h Integration: connect LLM pipeline to extension, wire up diff → classify → explain → quiz flow All together
10–14 4h Live preview panel (iframe-based), Bug Dashboard (Recharts in webview) Dev 1+2: preview, Dev 3+4: dashboard
14–18 4h End-to-end demo flow: plant bugs in demo app, test full pipeline, fix edge cases All together
18–21 3h Polish: animations, loading states, error handling, seed dashboard data Dev 1+2: polish, Dev 3+4: seed data
21–23 2h Demo prep: script the walkthrough, rehearse 3+ times, pre-record backup video All together
23–24 1h Buffer for fires, final testing, sleep deprivation management All together

Prompt Engineering Notes

Bug Classification Prompt (example)

You are a coding education assistant. Given a code diff (before and after a bug fix), your job is to:

1. **Classify** the bug into exactly one category:
   - CSS Layout / Specificity
   - Null / Undefined Reference
   - Off-by-one Error
   - Async / Race Condition
   - State Management
   - Type Mismatch
   - Logic Error
   - Syntax Error

2. **Explain** why the original code was broken in 2-3 sentences. Write for a student learning to code. Use plain English. Reference the specific lines that changed.

3. **Name the concept** — what CS/programming concept does this bug relate to? (e.g., "Immutability", "Event Loop", "CSS Cascade")

4. **Generate a quiz question** — one multiple-choice question that tests whether the student understands WHY the bug occurred (not just WHAT the fix was).

Respond in JSON format:
{
  "category": "...",
  "explanation": "...",
  "concept": "...",
  "quiz": {
    "question": "...",
    "options": ["A) ...", "B) ...", "C) ...", "D) ..."],
    "correct": "B",
    "explanation": "..."
  }
}

Why This Wins

  1. Timely narrative — "anti-vibe-coding" is the conversation happening RIGHT NOW in tech education. Judges will have opinions on this. Your project takes a clear, positive stance.
  2. It's a real product, not a toy — VS Code extension with persistent bug tracking feels like something students would actually install and use.
  3. The dashboard is the money shot — showing a visualization of a student's coding weaknesses over time is the moment judges lean forward. It transforms a bug fixer into a learning analytics platform.
  4. Education theme fit is perfect — every feature maps to pedagogy: active recall (quiz), spaced awareness (dashboard), concept linking (explanations).
  5. Technical depth is real — VS Code extension API + LLM pipeline + live preview + data visualization. This is not a ChatGPT wrapper.
  6. The demo is self-contained — you break something, you fix it, you learn from it. The entire loop happens in 2 minutes on one screen.

Comparison: VisualDebugger vs Other Top Picks

VisualDebugger FormBreaker DocTranslate
Theme Education Re-engineering Healthcare
Demo Impact 5 — live bug fix + dashboard 5 — side-by-side PDF fill 4 — before/after translation
Feasibility 3–4 — VS Code extension is tricky 5 — pre-select forms 5 — clean LLM pipeline
Judge Appeal 5 — timely, every dev relates 5 — universal pain 5 — health equity
Uniqueness 5 — nobody else will build this 3 — civic tech is known 3 — translation tools exist
Risk Medium — extension dev complexity Low Low
Ceiling if it works Highest — "I want this" reaction High High

Previous Research: All 9 Ideas (Archived)

Click to expand full research on all 9 original ideas

Education Ideas

E1: GradeGhost (Safe and Polished)

  • One-liner: AI grading co-pilot that takes a teacher's rubric and a stack of student essays, then produces draft feedback and scores that the teacher reviews and approves in a streamlined UI — cutting grading time by 80%.
  • The Pain: High school teachers grade 120–180 essays per assignment at 8–12 min each = 16–36 hours per cycle. #1 contributor to burnout (NEA). Students get shallow feedback ("Good job") because teachers are overwhelmed.
  • The Magic Moment: Split screen. Teacher uploads a rubric photo + 5 essays. System processes in ~30 sec. Grading queue appears — click Student 1, see highlighted passages, rubric criteria with suggested scores (e.g., "Thesis Clarity: 4/5"), and a draft feedback paragraph. Teacher tweaks one score, edits one sentence, hits "Approve." 90 seconds vs 10 minutes. Export to CSV + feedback PDFs.
  • Tech Stack: Next.js 14 + shadcn/ui + Tailwind, Claude API, GPT-4o Vision, Mammoth.js, pdf.js, Supabase, jsPDF + PapaParse, Vercel

E2: RewindEd (Ambitious and Impressive)

  • One-liner: Records a lecture, then auto-generates a personalized, interactive "replay" for each student — highlighting parts they need to re-hear based on knowledge gaps, with inline micro-quizzes.
  • The Pain: Lecture recordings are the #1 study resource but the worst format for learning. Passive re-watching is no better than not reviewing at all.
  • Tech Stack: Next.js 14, Video.js, OpenAI Whisper, Gemini 1.5 Flash, ChromaDB, Supabase, Vercel + Railway

E3: SyllabusNegotiator (Creative Wildcard)

  • One-liner: Students collaboratively negotiate their course syllabus with their professor by proposing, debating, and voting on topics and assessments — AI synthesizes into a class-authored syllabus.
  • The Pain: Student agency predicts engagement, yet syllabi are unilateral. Negotiated syllabi → 23% higher satisfaction but logistics are nightmarish.
  • Tech Stack: Next.js 14, Socket.io/Supabase Realtime, Claude API, Recharts, React-PDF, Supabase, Vercel

Healthcare Ideas

H1: DocTranslate (Safe and Polished)

  • One-liner: Takes medical discharge instructions and re-renders at 5th-grade reading level in patient's preferred language, with visual aids and read-aloud.
  • The Pain: 88 million US adults have limited health literacy. Poor comprehension → 30% higher readmission.
  • Tech Stack: Next.js 14, Claude/GPT-4o, RxNorm, openFDA, ElevenLabs, Synthea, Vercel

H2: ChartWhisperer (Ambitious and Impressive)

  • One-liner: Ambient AI scribe that generates SOAP notes in real-time from doctor-patient conversations with inline citations.
  • The Pain: Physicians spend 1h50m on documentation per 1h of patient care. #1 burnout driver.
  • Tech Stack: Next.js 14, Vercel AI SDK, Whisper, Claude/GPT-4o, LangChain + Chroma, RxNorm + openFDA, Supabase

H3: KinLoop (Creative and High-Impact)

  • One-liner: Family caregiver coordination: discharge docs → shared task boards, medication cards, AI early warning from daily check-ins.
  • The Pain: 53 million unpaid family caregivers. Care plans scattered across portals and group texts.
  • Tech Stack: Next.js 14, @hello-pangea/dnd, Claude API, openFDA + RxNorm, Supabase, Recharts, Resend, Vercel

Re-engineering Ideas

R1: FormBreaker (Safe and Polished)

  • One-liner: Upload bureaucratic PDF → conversational wizard fills it out with live side-by-side PDF population.
  • The Pain: 6.5 billion hours/year on federal tax compliance. Forms use jargon and conditional logic.
  • Tech Stack: Next.js 14, pdf.js, PyMuPDF, Marker, GPT-4o/Claude, pdf-lib, Vercel + Railway

R2: DeadCode Detective (Balanced)

  • One-liner: GitHub URL → interactive codebase autopsy with dead code map, tribal knowledge recovery, modernization roadmap.
  • The Pain: 42% of dev time on maintenance/tech debt. 25–30% dead code in mature codebases.
  • Tech Stack: Next.js, React Flow, Tree-sitter, simple-git, Claude/GPT-4o, FastAPI, Recharts, Vercel + Railway

R3: Re:Flow (Ambitious)

  • One-liner: Record a repetitive workflow → AI reverse-engineers into editable, executable automation.
  • The Pain: 58% of workday on "work about work." RPA tools cost $420+/month, 50% fail in year one.
  • Tech Stack: Next.js, React Flow, Chrome Extension, GPT-4o/Claude, PyMuPDF, Slack webhook, Vercel

Original Comparison Table

Idea Demo Impact Feasibility Judge Appeal Data Access Uniqueness TOTAL
E1 GradeGhost 4 5 4 5 3 21
E2 RewindEd 5 3 4 4 5 21
E3 SyllabusNegotiator 4 4 4 5 5 22
H1 DocTranslate 4 5 5 5 3 22
H2 ChartWhisperer 5 3 5 4 3 20
H3 KinLoop 4 3 5 4 4 20
R1 FormBreaker 5 5 5 5 3 23
R2 DeadCode Detective 5 3 4 5 4 21
R3 Re:Flow 5 2 5 3 5 20

Key Principles

  1. Demo-driven development — build the demo flow first. Everything the judge won't see doesn't exist.
  2. Prompt engineering is your ML — well-crafted prompts with structured output (JSON mode) get you 90% there.
  3. Seed data is not cheating — pre-populate with realistic data. Every winning team does this.
  4. One "wow" moment — the Bug Dashboard is yours. Practice hitting it cleanly.
  5. Deploy early — have the extension installable from VSIX by hour 14. Nothing kills a demo like "it works on my machine."