Mallik Cheripally
All work

Learning Hub

A training platform I built end to end, from product plan to grading engine

Learning Hub trains people to annotate data to production quality. Learners practice inside the real editor and get feedback that shows, on the image or in the 3D point cloud, exactly how their work differs from the reference. I built it from the ground up: the product plan, the design, the frontend, the backend, and the engines that compare and grade. It's in active use across every part of it.

One car in a street scan under review: a white reference box around it, the learner's red box overshooting it, the part of the car left outside that box lit gold, and the rest of the capture carrying on behind, parked cars in their own labels and the buildings past them

My part

  • Product and feature planning
  • UX and interface design
  • Entire frontend and backend
  • Comparison and grading engines

At a glance

Built
From the ground up
Engines
Comparison and grading, both new
Label types
2D and 3D, including point clouds
Built for
Learners, course authors, team leads
Practice
Inside the real annotation editor

The product

A course is a run of lessons: rich text, video, quizzes and practice. Leads assign courses with due dates and mark which ones are required. Quizzes draw from pools of questions and answers, so a retry tests the idea again rather than your memory of the last attempt.

Practice is where it gets serious. A practice lesson opens the full production editor with a dataset and a brief: what to label, how, with which categories, or what to review. When the learner finishes, Learning Hub compares their work with a reference set and shows them the difference, visually and in numbers. They can go again as many times as they need.

Feedback you can see

A score tells a learner they missed. The feedback shows where, in which direction, and by how much.

  • Wrong category

    Label a truck as a car and it's flagged as a major mistake, visually and in the result.

  • Overshoot

    Where a box covers more than the object, that extra area is drawn in its own color and texture.

  • Undershoot

    The part of the object a box missed is drawn differently, so excess and shortfall never read as the same error.

  • The numbers

    Pixel deviation on every edge, area overlap and IoU, for anyone who wants the precise picture.

A car in a lidar scan seen side on, a white reference box around it and a red learner's box overshooting the roof and tail, the uncovered nose lit gold
Reference box in white, the learner's box in red. Red marks the volume their box claims and the car does not fill. Gold marks the part of the car their box left out.

Feedback in 3D

Boxes on an image were the start. The same feedback had to work in lidar point clouds, where a label is a 3D box or a 3D polygon, and a learner can look at it from any angle.

For 3D boxes, extra boxes show the missed and overshot volume. 3D polygons were the hardest problem in the whole system. I carried the same visual language into the point cloud, coloring points and regions by the kind of error, so a learner can orbit the scene and see precisely what they got wrong.

The grading engine

None of this existed before. I built a comparison engine for every 2D and 3D label type, then a grading engine on top of it, with rules I defined from years of seeing how production annotation is actually reviewed. It grades:

  • Geometric quality and accuracy
  • Wrong label categories
  • Missing labels: objects that should have been labeled and weren't
  • Extra labels that shouldn't exist
  • Errors in label, frame and dataset attributes
  • Minor versus major mistakes, and how much each one counts

Courses can end in a final exam on a real dataset. The learner submits, the grader scores it automatically, and it passes or fails against thresholds and rules the course author sets, with as many attempts as the course allows.

For course authors

Team and project leads build their own training in Creator Studio, for a label type or a specific project, without waiting on anyone. I built that too, along with the media side behind it.

  • Lessons

    Rich, visual text lessons, uploaded content, and video that's uploaded and processed for playback.

  • Practice

    Practice lessons and feedback activities, configured against any dataset.

  • Exams

    Curated exam datasets, pass and fail thresholds, and grading rules.

  • Quizzes

    Pools of questions and answer choices, so every attempt is different.

  • Media Studio

    A shared pool of images per course, uploaded once and reused anywhere in it.

  • Assignments

    Courses assigned to learners with due dates and a required flag.

For team leads

  • Reporting

    See how annotators perform across courses and which label types each is strongest in, and combine results from several courses to judge who fits a project.

  • Progress tracking

    See who is working through which course, who's stuck, and which unfinished courses are close to their due date.