AI learning product · 2026

LabLens

An effective tutor should diagnose the misconception first, then offer the smallest useful intervention.

My role

I designed and built the interactive lab experience, misconception checks, coaching route, fallback behavior and teacher-facing session summary.

Live product Demo video View source

Responsibility: Product, learning flow + implementation
Core stack: Next.js · React · Qwen 2.5-72B
Current scope: 3 deployed interactive mini-labs

The operating problem

Students often receive explanations before anyone checks why their mental model is failing. LabLens turns that diagnostic moment into the center of the learning flow.

Evidence

What exists in the current build.

These numbers describe tested product scope—not production impact.

3interactive mini-labs, deployed

72Bparameter model in the coaching route

3hand-authored misconception maps

0lessons broken when the model is down

Inside the build

Screenshots of the running product.

Captured from the application itself—no mockups, no rendered concepts.

LabLens replying to a misconception with a water-pipe analogy question instead of the answer — **A real Socratic exchange**The learner says current gets used up by R1. LabLens does not correct them — it asks about water flowing through two connected pipes, and tags the misconception it detected.

LabLens circuits mini-lab before submission, with the problem statement and reasoning composer — **Predict before you measure**Every lab asks for a commitment and the reasoning behind it before the experiment runs. That answer is the diagnostic signal.

LabLens technology page showing the request path from student input through the API route to the model — **The route, documented in-product**The technology page states the actual path — student input, Next.js route, Featherless, Qwen2.5-72B — and the structured JSON it returns.

System workflow

From input to accountable action.

The interface keeps model work, deterministic checks and human responsibility distinct.

Predict

Ask for a commitment before the experiment.

The learner predicts an outcome and briefly explains their reasoning before seeing the result.

Evidence

The response creates a diagnostic signal rather than a passive completion event.

01 / Learning design

The interaction is the evidence.

The product does not treat time-on-page as understanding. It captures a prediction, an observed experiment and a follow-up explanation.

This creates a compact trail a teacher can inspect without turning the tutor into an automated grader.

02 / Resilience

The core lesson should not depend on a model call.

Each lab includes a bounded offline path so the activity remains useful when an external model is unavailable.

AI improves the specificity of the coaching prompt; it is not allowed to become the only source of instructional value.

Design principles

Diagnose before explaining.
Prefer one useful hint to a complete answer.
Make reasoning visible to the learner and teacher.
Keep the core learning experience resilient.

Next case study

CrewVerify