A focused empty state avoids a blank dashboard and guides the learner toward the first useful action.
Role
Product Designer
The project
Project overview
Calibrate is an AI-assisted study concept designed around a simple problem: recognizing information is not the same as being able to explain it.
The app turns imported notes into editable, source-linked review cards. Before answering, learners rate how confident they feel. They then explain the concept from memory, receive feedback tied to their own answer, and see which concepts should be reviewed next.
The central product hypothesis is that comparing perceived confidence with explanation quality can reveal false mastery and make review sessions more focused.
Problem
Learners can feel familiar with a concept while still being unable to explain it without support. Many study tools reinforce repetition and recognition, but provide limited insight into the reasoning or concepts that are still missing.
Goal
Design a mobile study experience that helps learners test what they can retrieve, understand why an explanation is incomplete, and decide what to review next.
My role
I defined the product concept, designed the core learning loop, mapped the mobile experience, and explored how AI-generated feedback could remain understandable, traceable, and grounded in the learner’s imported material.
Scope
Product framing, learning-science research synthesis, user flows, wireframes, AI interaction design, UI design, prototyping, and design exploration.
Defining the product
Product framing
Primary user
Higher-education learners studying subjects where understanding mechanisms and relationships matters more than memorizing isolated definitions.
Core job to be done
Help me identify what I cannot yet explain and decide what deserves attention in my next review session.
MVP scope
Import study material, generate editable source-linked cards, capture confidence, collect an open answer, provide traceable feedback, assign a learning state, and recommend the next review.
Main product risks
Unreliable AI evaluation, feedback that appears more certain than it is, privacy of imported material, and excessive cognitive load during review.
Understanding the problem
Research and learning principles
To frame the concept, I synthesised learning principles related to active recall, retrieval practice, explanation, and confidence calibration.
This phase was used to define product hypotheses rather than claim validated learning outcomes. The main hypothesis was that learners need more than repeated exposure: they need to compare what they believe they know with what they can retrieve and explain.
Three design hypotheses
Familiarity can feel like mastery
Seeing familiar material can create confidence without proving that the concept can be retrieved independently.
Design response:
Ask learners to explain before revealing the source or expected concepts.
Binary feedback is not enough
A right-or-wrong result does not show which part of an explanation was accurate, vague, incomplete, or missing.
Design response:
Connect feedback to specific segments of the learner’s own answer.
Repetition needs prioritization
Knowing that a card was difficult does not automatically tell the learner what should be reviewed next.
Design response:
Convert each answer into a learning state that influences the next review session.
Behavioral learning profiles
Instead of building demographic personas, I focused on learning behaviors.
The goal was to understand how different learners revise, where false confidence appears, and what kind of feedback helps them improve.
Competitive audit
I analyzed common study and learning apps to understand how they support repetition, recognition, and engagement.
The audit showed that most tools are good at helping learners review more often, but less effective at revealing whether a learner can explain a concept without support.
Most tools help learners review more often, but few reveal whether they can explain a concept without support.
This gap shaped Calibrate’s core loop: confidence, open recall, explainable feedback, and review priority.
From recognition to calibrated recall
The product logic
The core idea behind Calibrate is that understanding should not be measured by recognition alone.
A learner may recognize a concept and still struggle to explain it without support.
Calibrate turns each review card into a calibration loop: the learner rates their confidence, explains the answer, and receives feedback based on what was actually retrieved.
1
Confidence before feedback
Confidence is captured before the learner sees the source, expected concepts, or AI diagnosis.
2
Explanation as evidence
The learner answers in their own words, giving the system evidence about retrieved ideas, vague reasoning, and missing concepts.
3
Diagnosis becomes action
The answer is translated into a learning state that helps determine what should be reviewed next.
Translating the logic into an experience
Key UX decisions
After defining the logic, I focused on a few decisions that made the experience both cognitively useful and simple enough for mobile use.
Each choice was designed to reduce passive studying, protect the recall process, and turn AI feedback into a clear next action.
01
Reach the first useful action quickly
New users are sent directly to the first useful action: importing study material. Because Calibrate only becomes valuable once personal notes are added, the first screen focuses on creating the first review set instead of stretching the entry flow.
02
Protect effortful recall
During the answer step, the interface hides the source, expected concepts, and the card’s learning state.
Learners respond in their own words through voice or text, preserving the same open-recall task without relying on predefined answers.
03
Make AI feedback inspectable
A single score may be easy to scan, but it does not explain how an answer should improve.
Calibrate connects feedback to highlighted segments of the learner’s wording, making it possible to inspect what was strong, vague, incomplete, or missing.
Global verdict
Highlighted evidence
Focused explanation
04
Turn diagnosis into the next action
After each answer, Calibrate assigns a learning state: Blind Spot, Fragile, Solid, or Mastered.
These states shape the recommended Smart Review, while learners can still adjust the focus and session length according to their immediate goal.
Learning states
Card detail
Smart Review
The complete experience
End-to-end learning loop
After defining the product logic, I mapped the full learning loop: from imported material to a prioritized next review session.
The goal was to show how each interaction supports the same principle: learners should not only review cards, but understand what they can actually explain.
Learners add a PDF, photos, or pasted text and organise the material into one or more review sets.
A focused empty state directs the learner towards the first useful action.
Different source formats can be combined within the same import.
02
Generate and inspect source-linked cards
AI extracts key concepts and generates review cards linked to the imported material. Before studying, learners can inspect and edit the generated set.
Calibrate shows the generation process and produces editable review cards linked to the imported material.
The processing state explains what the system is doing.
Generated cards can be inspected and edited before review.
03
Rate confidence and explain
Before answering, learners estimate how ready they feel to explain the concept. They then respond by voice or text without seeing the source.
Before answering, learners estimate how ready they feel to explain the concept. They then respond by voice or text without seeing the source. This creates a signal that can reveal possible mismatches between confidence and answer quality.
Confidence is captured before feedback can influence it.
Voice and text preserve the same open-recall task.
04
Inspect the diagnosis
The answer is compared with the expected concepts from the imported material. The first feedback layer provides a short verdict and highlights relevant parts of the learner’s explanation. Learners can open individual segments to understand why a phrase was considered strong, vague, incomplete, or missing.
The response is compared with the expected concepts from the imported material. Feedback begins with a short overview and can be inspected through highlighted answer segments.
The first layer provides a quick overview.
Highlighted segments connect feedback to the learner’s own words.
Focused explanations show what could be improved.
** WCAG AA check + checkmark '✓' symbol for accessibility
05
Prioritize the next review
After the session, cards are grouped by learning state. Learners can follow the recommended Smart Review, inspect an individual card’s history, or immediately retry a weak concept.
After the session, cards are grouped by learning state. Learners can follow the recommended Smart Review, inspect an individual card, or retry a weak concept.
Learning states make weak areas easier to locate.
Card details combine the answer, source, and diagnosis.
Learners retain control over the next session.
Design exploration
Designing the AI review
I explored three feedback models and evaluated them against three design criteria: scannability, explainability, and actionability. This was a design evaluation, not a user-testing result. The objective was to understand what each model communicated well and where it reduced the quality of the learning signal.
1
Checklist feedback
This direction separated the answer into covered, partial, and missing ideas. It was precise and auditable, but fragmented the feedback without clearly communicating the main issue in the explanation.
2
Percentage-based scoring
This direction translated the answer into a comprehension percentage supported by learned, partial, and missing concepts. It was fast to scan, but compressed a nuanced explanation into a number that could hide the reasoning mistake that needed attention.
3
Micro-synthesis
This direction used a short verdict to summarise the main issue in the answer. It created a clearer hierarchy, but the verdict was not sufficient on its own. Visible evidence was still needed to understand why the conclusion was reached.
Final direction
The selected direction combines a short verdict, highlighted answer segments, and focused explanations. The verdict communicates the main learning signal quickly. The highlighted response makes the diagnosis traceable to the learner’s own words, while focused explanations show what was strong, vague, incomplete, or missing. This direction does not remove uncertainty from AI evaluation, but it makes the system’s reasoning easier to inspect and question.
Final reflection
Prototype outcome
What the prototype demonstrates and what still needs validation.
Calibrate demonstrates an end-to-end learning concept that connects imported material, confidence, open recall, answer-level feedback, learning states, and review prioritisation within one continuous experience.
Its main contribution is not an automated score, but a system that helps learners inspect possible gaps between perceived confidence and the concepts they were able to explain.
What the prototype demonstrates
A coherent flow from imported notes to a prioritised review.
A way to compare confidence with open-answer quality.
Feedback connected to the learner’s answer and source material.
A learning-state system that turns diagnosis into a next action.
What remains to be validated
Whether learners use the confidence input consistently.
Whether highlighted feedback feels clear and appropriately cautious.
Whether learning states help learners choose what to review next.
How reliably the AI handles alternative valid explanations.
Whether the complete flow remains manageable during longer sessions.
How privacy and data retention should be communicated.
Next experiment
The next step would be a moderated concept and usability test with learners studying concept-heavy subjects.
The test would examine whether participants understand the confidence step, can interpret the diagnosis, can explain why a card received a specific learning state, and can choose an appropriate next review without assistance.
How success would be measured
The core hypothesis would be tested by tracking a calibration gap: the difference between a learner’s stated confidence before answering and the quality score their explanation actually receives. If Calibrate works, this gap should narrow across sessions as learners get better at judging what they can explain — not just what they recognize.
Key learnings
What this project helped me clarify as a Product Designer.
01
AI feedback is an interface
AI feedback should be designed as an inspectable interface, not presented as an unquestionable automated result.
02
Depth must remain usable
The challenge was to provide enough evidence to support reflection without turning every answer into a dense report.
03
The complete loop matters
The product value comes from connecting input, recall, diagnosis, prioritisation, and the next action rather than treating AI generation as an isolated feature.