ParikshaMitra

What it does# ParikshaMitra (परीक्षा मित्र)

About the Project

ParikshaMitra is an agentic AI study companion for Indian competitive exams. We built it around a simple belief: students do not just need more content—they need a system that can understand how they study, identify what they are weak at, and continuously adapt their preparation plan. Instead of behaving like a static chatbot or a passive doubt-solving tool, ParikshaMitra acts like an intelligent companion that can plan, test, analyze, and adapt across study sessions. :contentReference[oaicite:1]{index=1}

What Inspired Us

The inspiration came from a very visible problem in Indian exam preparation: millions of students compete every year, but most preparation systems still follow a one-size-fits-all approach. Students are often overwhelmed by PDFs, videos, mock tests, and schedules that do not adjust to their pace, weaknesses, or progress. Diagnostic feedback is usually shallow, and after a poor test, most students still do not know why they underperformed or what exactly to do next. That gap inspired us to think beyond content delivery and build a system that could dynamically orchestrate learning itself. :contentReference[oaicite:2]{index=2}

We were especially motivated by the idea of democratizing elite preparation. A student in a Tier-3 city should be able to receive the same kind of personalized guidance that usually comes from expensive coaching ecosystems. That vision became the foundation of ParikshaMitra. :contentReference[oaicite:3]{index=3}

What We Built

We built ParikshaMitra as a multi-agent cognitive system for autonomous exam preparation. Its workflow follows a loop:

$$ \text{Plan} \rightarrow \text{Test} \rightarrow \text{Analyze} \rightarrow \text{Adapt} $$

This loop allows the system to generate study plans, assess performance, diagnose conceptual gaps, and revise the next plan automatically without waiting for repeated user prompting. :contentReference[oaicite:4]{index=4}

At the core of the project is a 4+1 cognitive architecture:

• Planner (Strategist): allocates study hours based on exam priorities and student weaknesses.
• QuizMaster (Evaluator): generates quizzes, adjusts difficulty, and grades responses.
• Analyst (Diagnostician): computes readiness and identifies hidden prerequisite gaps.
• Companion (Motivator): turns system outputs into supportive, bilingual nudges.
• Orchestrator Agent: coordinates all other agents and manages global state. :contentReference[oaicite:5]{index=5}

This architecture helped us move from a normal chatbot design to a more agentic system where specialized roles work together in a structured way.

How We Built It

We built the project using a modern agent and app stack:

• LangGraph for stateful agent workflows and orchestration
• Google Gemini 2.0 Flash as the LLM core
• Python 3.11 for backend logic
• Streamlit for rapid MVP interface development
• SQLite + Pydantic v2 for persistent student history and structured state handling :contentReference[oaicite:6]{index=6}

The intelligence layer includes several mechanisms that make the system adaptive instead of static:

1. Spaced Repetition

We used an SM-2 style spaced repetition approach to decide when a topic should be reviewed again. Performance influences review intervals:

• High score: extend interval
• Medium score: maintain interval
• Low score: reset review sooner

This helps optimize revision timing instead of relying on fixed schedules. :contentReference[oaicite:7]{index=7}

2. Adaptive Difficulty

We designed a lightweight CAT-like difficulty adjustment system that changes question level in real time based on recent student performance. In simple terms:

$$ d_{t+1} = \begin{cases} d_t + 1, & \text{after sustained correct answers} \ d_t - 1, & \text{after repeated incorrect answers} \end{cases} $$

This keeps the student in a more productive challenge zone instead of making every learner attempt the same question set. :contentReference[oaicite:8]{index=8}

3. Concept Dependency Tracking

We modeled prerequisite relationships through a concept dependency DAG, so when a learner struggles on an advanced topic, the system can detect whether the actual problem lies in an earlier missing concept. This makes remediation much smarter than just saying “practice more.” :contentReference[oaicite:9]{index=9}

4. Exam Readiness Score

We also introduced a readiness score to quantify student confidence and preparation using multiple signals:

$$ R = 0.4(\text{Coverage}) + 0.3(\text{Mastery}) + 0.2(\text{Revision}) + 0.1(\text{Mock Trend}) $$

This gives a more holistic picture than raw marks alone. :contentReference[oaicite:10]{index=10}

MVP Features

Our hackathon MVP demonstrates the essential experience of the system:

• Smart onboarding with a short diagnostic
• Personalized study planning
• Dynamic quiz generation
• Performance dashboard
• Spaced repetition calendar
• Bilingual companion support in Hindi and English :contentReference[oaicite:11]{index=11}

One of the strongest moments in the user journey is the reflective replanning step: when a student struggles in a focused quiz, the system automatically revises the study schedule, increases time on weak chapters, and flags prerequisite topics for immediate review—without requiring a manual prompt. That is where the project most clearly demonstrates true agentic behavior. :contentReference[oaicite:12]{index=12}

What We Learned

Building ParikshaMitra taught us a lot about the difference between LLM apps and agentic systems.

First, we learned that good educational AI is not just about generating explanations. It is about maintaining state, tracking progress over time, and making decisions based on accumulated evidence. Persistent memory and orchestration turned out to be as important as the model itself. :contentReference[oaicite:13]{index=13}

Second, we learned that specialization improves reliability. Dividing the system into planner, evaluator, analyst, and companion roles made the workflow easier to design, debug, and reason about than forcing one monolithic agent to do everything. :contentReference[oaicite:14]{index=14}

Third, we realized that motivation matters in learning systems. A study companion should not only assess and optimize but also communicate in a way that keeps students engaged. That is why the bilingual companion layer became an important part of the experience. :contentReference[oaicite:15]{index=15}

Challenges We Faced

One major challenge was designing true autonomy rather than a chatbot that only reacts when asked. We had to think carefully about when the system should act on its own, when it should escalate, and how its internal state should flow between agents. Creating that orchestration logic was one of the hardest parts of the project. :contentReference[oaicite:16]{index=16}

Another challenge was balancing speed and intelligence for a hackathon MVP. We wanted meaningful adaptation, persistence, and explainability, but we also needed the system to remain fast and demo-friendly. This forced us to simplify some advanced ideas into lightweight but effective versions, such as CAT-lite difficulty adjustment and practical readiness scoring. :contentReference[oaicite:17]{index=17}

We also faced the challenge of converting educational logic into something computational. Concepts like mastery, readiness, prerequisite gaps, and revision timing sound intuitive in theory, but implementing them in a robust workflow required careful system design. :contentReference[oaicite:18]{index=18}

Why This Project Matters

ParikshaMitra is not just another AI study tool. It is an attempt to move from static EdTech to autonomous learning orchestration. Existing platforms often provide lectures, question banks, or explanations, but they do not continuously reason about what the student should do next. ParikshaMitra aims to close that gap by bringing agentic decision-making into exam preparation. :contentReference[oaicite:19]{index=19}

We believe this can create both educational and social impact:

• Better outcomes through targeted adaptation
• More efficient revision through intelligent scheduling
• Greater access to personalized preparation at very low cost
• Strong alignment with competency-based and personalized education goals :contentReference[oaicite:20]{index=20}

Vision Forward

Our long-term vision is to expand beyond the MVP into broader exam ecosystems such as NEET, UPSC, and other high-stakes Indian exams, while also adding richer interaction channels like voice and messaging integrations. We also see future potential in parent dashboards and institutional licensing for schools and coaching centers. :contentReference[oaicite:21]{index=21}

Final Reflection

ParikshaMitra began as a response to a simple but urgent question:

What if an AI could study the student, so the student could study smarter?

Through this project, we explored how agentic AI can make preparation more adaptive, affordable, and humane. We did not just build a quiz bot or a scheduler—we built the foundation of an intelligent study companion that learns from student performance and responds with purpose. That journey challenged us technically and conceptually, and it showed us how AI can be used not only to answer questions, but to guide learning itself. :contentReference[oaicite:22]{index=22}

Built With

• across
• adaptive
• adjustment)
• algorithm)
• analyst
• api-integration
• companion
• concept
• dependency
• difficulty
• graph
• memory
• orchestrator)
• persistent
• prerequisite
• quizmaster
• repetition
• sessions
• spaced
• testing
• tracking)