Inspiration

Most AI tutors are very good at answering questions. But answering a student's question is not necessarily the same as helping that student learn.

A student may ask:

"I don't understand terminal velocity."

The real problem may not be terminal velocity at all. The student may have misunderstood viscosity, Stokes' law, force balance, or even the meaning of net force.

That observation inspired "eduPhysics Physics Mastery Agent".

We wanted to build an AI agent that does something fundamentally different from a conventional question-and-answer tutor:

"Instead of only answering what a student asks, the agent tries to discover what the student actually needs to learn next."

The vision is to combine Gemini-powered reasoning with the educational resources of the eduPhysics ecosystem to create an adaptive Physics learning companion that can diagnose misconceptions, select an appropriate teaching strategy, challenge the learner, evaluate reasoning, and continuously guide the student toward mastery.


What it does

The eduPhysics Physics Mastery Agent is designed around an agentic learning loop:

"Diagnose → Retrieve → Explain → Challenge → Evaluate → Detect → Adapt → Re-test → Advance"

A student can begin with a simple statement such as:

"I don't understand terminal velocity."

Rather than immediately producing a long explanation, the agent can ask targeted diagnostic questions.

It attempts to determine whether the difficulty comes from:

  • a conceptual misunderstanding,
  • a missing prerequisite,
  • incorrect use of a formula,
  • mathematical reasoning,
  • interpretation of forces,
  • or an inability to connect the concept to a physical situation.

The agent can then retrieve relevant educational material, explain the concept at an appropriate level, generate a targeted challenge, and evaluate the student's response.

The most important part of the workflow is misconception detection.

For example, if a student says:

"Gravity stops acting when terminal velocity is reached."

the agent should recognize that the student has confused "zero net force" with "zero gravitational force.

Instead of simply marking the answer wrong, the agent can identify the misconception, explain the force balance, and generate a new question specifically designed to test whether the misconception has been corrected.

The goal is to create a learning loop rather than a one-time answer.


How we built it

The project is being designed as an agentic AI system rather than a single chatbot prompt.

At the center is the "Physics Mastery Agent", which uses Gemini to reason about the student's current learning state and decide what action should happen next.

The agent is organized around several capabilities:

  1. "Student diagnosis"
    Identifying conceptual and prerequisite weaknesses.

  2. "Physics knowledge retrieval" Finding relevant material from the eduPhysics learning ecosystem.

  3. "Adaptive explanation"
    Selecting an explanation appropriate to the student's demonstrated difficulty.

  4. "Challenge generation" Creating targeted conceptual and numerical questions rather than random practice.

  5. "Reasoning evaluation"
    Looking beyond the final answer to examine the student's approach, equations, assumptions, units, and reasoning.

  6. "Misconception detection"
    Identifying the underlying reason for an incorrect response.

  7. "Adaptive intervention" Deciding whether to explain again, simplify, go back to a prerequisite, provide another example, or increase difficulty.

  8. "Mastery progression"
    Determining when the student is ready to move to the next concept.

The architecture is intentionally designed so that the agent can make decisions about the "next learning action", rather than following a fixed sequence of questions.

The initial demonstration focuses on a Physics concept such as terminal velocity because it provides a particularly clear example of misconception-driven learning.


Challenges we ran into

One of the biggest challenges was defining what makes an educational AI system genuinely "agentic".

Simply connecting Gemini to a Physics question-answering interface would not be enough.

We therefore focused on the decision loop:

"What does the student know?" ↓ "What is the likely root misconception?" ↓ "What should the agent do next?" ↓ "Did the intervention work?" ↓ "What should happen after the student's next response?"

Another challenge is balancing helpful explanations with cognitive overload.

A Physics student does not always need more information. Sometimes the student needs one carefully selected question that exposes the missing concept.

A second challenge is preventing the system from confidently presenting an incorrect Physics explanation. This makes grounded retrieval and careful evaluation particularly important.

A third challenge is designing the system so that it can serve different learners—from a student learning a Class 11 concept for the first time to a student preparing for competitive examinations.


Accomplishments that we're proud of

We are building the project around a simple but powerful principle:

"The next best learning action should depend on the student's demonstrated understanding."

We are particularly proud of the misconception-repair workflow.

Instead of:

"Question → AI answer → End"

the system is designed to create:

"Question → Diagnosis → Targeted intervention → Student reasoning → Evaluation → Adaptation → Mastery"

We are also proud of the opportunity to connect an agentic learning system with the broader eduPhysics ecosystem, including Physics learning resources, conceptual explanations, practice questions, previous-year questions, practical learning resources, and topic-based Master Hubs.

This creates the possibility of turning a large educational content library into an intelligent learning pathway rather than simply a collection of web pages.


What we learned

Building an educational agent taught us that "knowing the answer and knowing how to teach the answer are two different problems".

A strong educational agent must reason about the learner, not just the subject.

We also learned that agentic systems become more useful when their actions are explicit and measurable.

For Physics education, this means distinguishing between:

  • knowing a formula,
  • understanding a concept,
  • applying a principle,
  • solving a numerical problem,
  • interpreting a physical situation,
  • and explaining the reasoning.

We learned that a useful AI tutor should be willing to step backward.

If a student struggles with terminal velocity, the correct next step may be viscosity or force balance—not another terminal-velocity explanation.

That ability to navigate backward through prerequisites and forward toward mastery is central to our vision.


What's next for eduPhysics Physics Mastery Agent

Our next goal is to expand the agent from a focused demonstration into a broader Physics mastery system.

We envision a personalized "Physics Mastery Map" in which students can see their understanding across concepts and identify the areas that need attention.

Future capabilities include:

  • prerequisite-aware learning paths,
  • adaptive numerical problem solving,
  • CBSE-focused preparation,
  • NEET and JEE-oriented practice,
  • Physics practical and viva preparation,
  • visual and experiment-based explanations,
  • personalized revision plans,
  • misconception tracking,
  • mastery-based progression,
  • and deeper integration with the eduPhysics Master Hubs.

Ultimately, our goal is not to build another AI that simply answers Physics questions.

We want to build an agent that can continuously ask:

"What does this student understand, what are they missing, and what is the best thing I can do next to help them master Physics?"

“Instead of only answering what a student asks, the agent tries to discover what the student actually needs to learn next.”

That is the vision behind the "eduPhysics Physics Mastery Agent".

Built With

Share this project:

Updates

posted an update

This update is about the moment that convinced us this project was worth building: watching the agent catch a misconception a normal AI tutor would sail right past.

What's new

Diagnostic-first flow — the agent now teaches a concept, then asks a targeted diagnostic question before moving to practice, rather than assuming understanding just because a student didn't ask a follow-up. Misconception detection — when a student answers "the net force is downward, gravity is still acting" at terminal velocity, the agent identifies the specific confusion (zero net force vs. zero gravitational force) instead of just marking it wrong. Persistent mastery state — every interaction now updates a per-topic mastery score (Proficient / Developing / Needs Review) that carries across sessions, not just within one conversation. Structured numerical problem-solving — quantitative problems now walk through units → formula → substitution → calculation → final answer, matching the format examiners expect.

What's next

Expanding the diagnostic loop beyond terminal velocity to the full mechanics unit, and turning the Mastery Map mockup above into a real, persistent student-facing dashboard.

Live demo: https://eduphysics-physics-mastery-agent-poqvtgavdq-el.a.run.app Code: https://github.com/Rajansmoorthy/eduphysics-physics-mastery-agent

Log in or sign up for Devpost to join the conversation.

posted an update

eduPhysics Physics Mastery Agent — From Prototype to Cloud Deployment Update

A major milestone for the eduPhysics Physics Mastery Agent!

The project has evolved from an initial Physics tutoring prototype into a tested, containerized, cloud-deployed agentic learning system.

The agent is designed around an adaptive learning loop:

Diagnose → Retrieve → Explain → Challenge → Evaluate → Detect → Adapt → Re-test → Advance

Instead of simply answering a Physics question, the agent attempts to understand what the student actually knows and determine the next best learning action.

What has been implemented

• Evidence-grounded Physics knowledge retrieval • Student-response evaluation • Persistent mastery state across learning turns • Adaptive tutoring behavior • Misconception-aware learning workflow • Diagnostic questioning • Adaptive conceptual and numerical challenges • Student reasoning evaluation • Student-readable mathematical solutions without raw LaTeX formatting • Physics corpus packaged into the production container • Docker containerization • Google Cloud deployment

Testing

The project currently has 121 automated tests passing.

This was particularly important because the goal is not just to demonstrate an AI response, but to verify that the agent's learning state and tutoring behavior persist across interactions.

Cloud deployment

The tested container has now been deployed to Google Cloud Run in asia-south1, with Gemini-powered inference and the required cloud infrastructure configured.

Example learning interaction

When a student says:

"I don't understand terminal velocity."

the agent doesn't necessarily start with a long explanation.

It can first investigate the student's understanding of gravity, air resistance, force balance, and acceleration.

If the student demonstrates partial understanding, the agent can move forward to questions such as:

"What happens when the upward drag force becomes equal to the downward weight?"

This allows the system to turn a question-answer interaction into a learning progression.

The bigger idea

The central principle behind the project is:

Don't just answer the student's question. Determine what the student needs to learn next.

The next phase is to expand this focused demonstration into a broader Physics Mastery Map, with prerequisite-aware learning paths, adaptive numerical practice, CBSE/NEET/JEE preparation, misconception tracking, revision planning, and deeper integration with the eduPhysics learning ecosystem.

Building this as a solo project has been challenging, but seeing the agent move from local experiments → automated testing → Docker → Google Cloud deployment has been a major milestone.

The journey from AI tutor to AI learning companion continues.

Log in or sign up for Devpost to join the conversation.

Submission history