A Socratic STEM tutor that diagnoses misconceptions — not just answers questions.

Understanding is not recall. LabLens guides students to the insight themselves; it never gives the answer first.

Students interact with live mini-labs, explain their thinking, and receive a targeted hint tied to the exact misconception detected — not a generic explanation. A follow-up check verifies whether the misconception was actually repaired, and a teacher-ready report summarizes the cognitive gap and next step.

STEM students don't fail because they lack answers — they fail because they hold the wrong mental model and never realize it. A student who believes "current gets used up across resistors" will keep making the same mistake no matter how many times they read the correct explanation. Traditional tutoring tools (ChatGPT, Wolfram Alpha, Khan Academy) detect the wrong answer and deliver the right answer. None of them pinpoint the underlying cognitive misconception and force the student to reason past it.

LabLens separates diagnosing from correcting:

• A live mini-lab lets students manipulate variables (circuit voltages, reaction ratios, function inputs) before typing anything — their lab choices already reveal their mental model.
• A Socratic AI engine (Qwen 2.5-72B via Featherless.ai) reads the student's explanation and their lab behavior, classifies the exact misconception, and issues a targeted hint — never the full answer.
• A follow-up check asks a structurally identical question. If the student answers correctly, the misconception is marked repaired. If not, the hint deepens by one level.
• A teacher report summarizes the diagnosed misconception, the hint chain used, and the recommended next step — exportable for classroom review.

The AI never lectures. It asks the question that forces the student to find the gap themselves.

  STUDENT INPUT                          SOCRATIC ENGINE (Featherless.ai)
  Mini-lab interaction                   Qwen/Qwen2.5-72B-Instruct
       │                                        │
       │  Live parameter controls               │  System prompt encodes:
       │  (voltage, moles, derivatives)         │  misconception taxonomy
       ▼                                        ▼  + Socratic hint ladder
 ┌─────────────────────────────────────────────────────────────────────────┐
 │  Next.js App Router   ·   /app/api/chat   ·   Vercel Edge               │
 │                                                                          │
 │  Request pipeline:                                                       │
 │   1. Student submits explanation + lab state                             │
 │   2. API route calls Featherless (OpenAI-compatible endpoint)            │
 │   3. Model returns: { misconception, hint, chips, next_move }            │
 │   4. UI renders targeted hint + cognitive progress chips                 │
 │   5. Follow-up check re-enters pipeline at deeper hint level             │
 └─────────────────────────────────────────────────────────────────────────┘
        │                          │                          │
 ┌──────┴──────┐          ┌────────┴───────┐         ┌───────┴──────────┐
 │  Physics     │          │  Calculus       │         │  Chemistry        │
 │  Series      │          │  Derivative     │         │  Limiting         │
 │  circuits    │          │  vs. value      │         │  reactants        │
 └─────────────┘          └────────────────┘         └──────────────────┘

More subjects (Biology, Statistics, Mechanics) are planned for the next sprint.

git clone https://github.com/Kimberly-Huang/Lablens.git
cd Lablens
npm install

Create .env.local:

FEATHERLESS_API_KEY=your_key_here

npm run dev
# → http://localhost:3000

If no API key is present, the app uses a deterministic fallback coach so the demo still runs offline.

1. Push to GitHub.
2. Import repo at vercel.com/new.
3. Add FEATHERLESS_API_KEY under Settings → Environment Variables.
4. Deploy — no other configuration needed.

• Three subjects only. The misconception taxonomy is hand-authored for Physics, Calculus, and Chemistry. Expanding to new subjects requires writing a new taxonomy entry and mini-lab component.
• Single-session memory. The hint chain resets on page refresh; there is no persistent student profile or learning history across sessions.
• No curriculum alignment. Misconceptions are mapped to concepts, not to specific standards (e.g., AP Physics curriculum units). A production version would tag each misconception to a standard code.
• AI confidence is opaque. The model returns a classification and hint but no confidence score; a low-confidence classification silently uses the same hint ladder as a high-confidence one.

• Persistent student profiles — track misconception history across sessions; flag recurring gaps for teacher review.
• Curriculum-aligned taxonomy — tag each misconception to AP / IB / Common Core codes so the teacher report maps directly to standards.
• Adaptive hint depth — use confidence scoring to decide whether to deepen the hint or try a different Socratic angle.
• More subjects — Biology (cell membrane transport), Statistics (p-value misinterpretation), Mechanics (Newton's Third Law symmetry).
• Classroom dashboard — aggregate misconception heatmaps across a class cohort in real time.