All work

PyxiScience · Assessments

One flow to create an assessment, and AI grading where the teacher keeps the final say

  • AI Product Design
  • 0 → 1 · Founding designer
  • 2025–2026
The new activity creation form, where the activity type is a single field, next to the class results dashboard with grade statistics and distribution charts

Overview

I merged three creation paths into a single flow, then designed AI-assisted grading where the teacher always makes the final decision.

Details

Product
Learning web app for higher-education teachers and students, at an AI edtech startup
Role
Founding Product Designer, sole designer
Team
6 people on product: CEO, Head of AI, CTO, developer, PM and me
Scope
Discovery · UX · UI · Coded prototypes

Problem

Three paths to the same thing

To create an assessment, teachers had to pick between three paths leading to three different objects that shared the same settings. It caused confusion, extra steps and lost time.

Solution

One flow, then a grading loop

A single creation flow with only the steps that matter, leading straight into AI grading that the teacher reviews, edits and chooses to share.

Result

Creation shipped, grading partly shipped

The new creation flow is in production and saves teachers 5 minutes per assessment. AI grading is partly live; the rest is shown here as a design vision.

Impact

3 1

creation paths, merged into a single flow shipped in production

–5 min

to create each assessment with the new flow

Context

Assessments are the core of the product: they feed the AI's recommendations for each student.

A painful creation flow means fewer assessments, so less data and less useful recommendations.

Our users are higher-education teachers: comfortable with digital tools without being experts. The interface had to stay readable with no training.

Before, reconstructed

Three tabs, three create buttons, each promising the same thing. Teachers didn't know which one to pick.

Goals

Make creating an assessment fast, then make AI grading something teachers can trust.

  1. Remove the upfront choice and the duplicated steps.
  2. Connect creation and grading as one loop.
  3. Let the AI do the heavy lifting without ever taking the decision away from the teacher.

Process

Three signals pointed to the same problem.

Teacher interviews

University teachers confirmed the frustration with creating assessments.

Amplitude

Usage data went further: in practice, only one of the three objects was really used.

Coded prototypes

I built testable prototypes with Claude Code to try interactions in real conditions, beyond a Figma prototype.

The CEO had also identified creation as a blocker.

Key insights

Insight 01

Three objects, one set of settings.

The three paths were almost identical, yet teachers had to choose one at the very start, without knowing which one fit.

Matrix of the settings of the three assessment objects: homework, exercise sheet and exam share 8 of their 11 settings (title, class, questions, due date, grading scale, answer key, paper or online, results sharing). Only hints, multiple attempts and time limit differ. Homework Exercise sheet Exam Title Class Questions Due date Grading scale Answer key Paper or online Results sharing Hints Multiple attempts Time limit 8 of 11 settings shared by all three objects

Insight 02

Only one of them was really used.

Amplitude showed that teachers had already picked a winner. The other two paths only added confusion.

Bar chart of assessments created per object: homework 82%, exam 12%, exercise sheet 6%. Share of assessments created, by object Homework 82% Exam 12% Exercise sheet 6%

Part 1

A single creation flow

Since the three objects shared the same settings and only one was really used, I proposed merging them into one path, with only the useful steps. Two decisions shaped it. One entry point: the type of activity is no longer a separate path but a choice inside the flow, so homework, exercise sheets and exams all start the same way. One exit point: every activity ends in the same grading and results, so students get the same kind of feedback whatever the format.

Creation is designed as the first half of one loop that leads straight into grading.

Before and after diagram. Before: three separate paths, homework, exercise sheets and exams, each with its own entry point, its own form and its own results, so three entries and three exits. After: one entry point, New activity, leads to a single flow (type, format, exercises) and ends in one place, grading and results, so students get the same feedback whatever the activity. BEFORE AFTER Homework tab + button Its own form Its own results Exercise sheets tab + button Its own form Its own results Exams tab + button Its own form Its own results 3 entry points · 3 exits, feedback differs by path New activity one entry point ONE FLOW Type homework, exam… Format online or on paper Exercises library or your own Grading & results one exit 1 entry point · 1 exit, the same feedback for every student

Shipped

Step 1, Details: the activity type (homework, exercise sheet, exam or other) is now a single field of the creation form Step 2, Format: online or on paper, then a bubble sheet or a mixed subject Step 3, Exercises: picking exercises from the platform library, with a summary and an answer sheet preview
One flow for every activity: the type that used to split creation into three paths is now a single field in the first step.

Part 2

AI grading, with the teacher in charge

The AI grades the paper, annotates it, detects errors and drafts a comment for the student. The teacher validates. That principle shaped the whole grading interface.

Shipped

Step 1: the teacher validates the structure of the subject, exercise by exercise, before grading starts Step 2: the teacher reviews class grades, with averages, pass rate and the grade distribution Step 3: once shared, each student sees their score, strengths and weaknesses
The teacher validates the structure, reviews grades, then chooses when students see their results.

Vision, not shipped

In wireframes: the AI matches papers to students and suggests criteria and a comment; the teacher adjusts, validates, then shares.

Design principles

The AI suggests, the teacher decides.

No grade is validated without the teacher. If they think the AI is wrong, they take over directly in the interface.

Make the AI's work readable.

Errors are annotated where they happen, so teachers can check quickly instead of rereading everything.

The teacher chooses what the student sees.

The grade and comment are only shared when the teacher decides.

Constraints

The main constraint was the shape in which the AI returned its results.

The AI's output wasn't always stable: the same paper could come back with a different grade or comment. Instead of hiding it, the interface flags those results so teachers review them before anything reaches students. Nothing is sent without a teacher's check.

Rather than designing everything in one go, we added features and steps progressively. Validating the structure, reviewing grades and sharing results are in production; AI-suggested criteria and comments on each paper are still a design vision.

Reflections

Ship even smaller.

I would release features in even smaller pieces, and let teachers try the tool earlier to improve from real use.

On an AI product, usage reveals trust.

Real use quickly shows where trust is won or lost, faster than any prototype.