TeacherAssistant Agent

Inspiration

As a university Computer Science student navigating rigorous semester workloads, I noticed a recurring friction point in higher education: the sheer administrative overhead instructors face. Designing a structured 16-week semester plan, drafting weekly lecture slides, generating robust assignments, quizzes, and rubrics, and then grading dozens—sometimes hundreds—of student submissions consumes countless hours that could otherwise be dedicated to direct student mentorship and research.

Furthermore, having previously built intelligent workflow prototypes like SafeRoute-AI (a web-based public transport safety and assistant architecture), I realized that Generative AI and automated backend agents could be leveraged not just for commuters, but to streamline academia. I wanted to build a comprehensive, reliable AI-based University Teacher Assistant System that acts as a co-pilot for educators—automating the heavy lifting of course creation while keeping human oversight firmly in the loop.

What it does

The system is architected as a modular, full-stack platform combining modern web technologies with robust AI orchestration to automate academic workflows:Course & Curriculum Generation Engine: Instructors input course objectives, and the system queries language models to generate a synchronized 16-week semester schedule. Each week automatically provisions:Lecture slide outlines and key concepts.Tailored assignments and quizzes.Detailed assessment rubrics.Submission & Assessment Pipeline: Students upload code or written submissions. The platform processes these inputs and prepares them for evaluation.AI-Assisted Grading & Approval Workflow: The AI evaluator analyzes student work against the generated rubrics, computing preliminary scores and feedback:$$\text{Final Grade} = \text{AI Evaluation} \times \text{Confidence Score} + \text{Instructor Adjustment}$$However, no grade is posted to the gradebook automatically; it enters an Instructor Review Queue where teachers can approve, edit, or reject the AI-suggested marks.LMS Export Module: Once approved, course materials and grades are bundled into LMS-ready formats for seamless migration into platforms like Moodle or Canvas. How we built it The project was built as a full-stack AI-powered application leveraging modern web frameworks and structured language model orchestration:

Frontend & Architecture: Designed a responsive, high-density instructor and student dashboard allowing seamless upload of assignments, real-time tracking of grading queues, and curriculum previews.

Backend & AI Orchestration: Integrated LLMs with strict JSON schema enforcement to ensure deterministic output generation for 16-week syllabi, lecture contents, and rubrics.

Asynchronous Processing: Implemented background task queues to handle bulk student submissions and AI grading evaluations asynchronously, ensuring zero UI lag.

Challenges we ran into Handling AI Hallucinations in Curricula: Early iterations occasionally produced inconsistent week-to-week topic progressions or overlapping quiz questions. We solved this by implementing a multi-step generation pipeline where the model first drafts a high-level course blueprint before expanding individual weekly modules.

Latency in Bulk Grading: Running AI evaluations for large batches of student submissions caused noticeable lag, which required implementing asynchronous task queues to process grading in the background.

Balancing Automation and Control: Finding the exact threshold for when an AI grade requires mandatory human review versus minor adjustments required fine-tuning the confidence scoring algorithm and designing an intuitive moderation interface for teachers.

Accomplishments that we're proud of Successfully creating an end-to-end co-pilot ecosystem that slashes hours of administrative friction for university instructors.

Implementing a foolproof Human-in-the-Loop (HITL) architecture that guarantees no AI-generated grade is published without mandatory human verification and approval.

Achieving reliable, structured multi-week output generation that transforms high-level course scopes into detailed, deployable weekly modules, slides, and assessment rubrics.

What we learned Human-in-the-Loop (HITL) Architecture: True automation in education cannot mean total autonomy; balancing speed with accountability through teacher approval workflows is essential.

Structured Generative Outputs: Generating 16-week curricula and weekly slide outlines requires strict prompt engineering and deterministic schema enforcement so outputs map cleanly into relational database structures.

LMS Interoperability: Understanding how Learning Management Systems ingest data highlighted the critical importance of standardized export schemas.

What's next for TeacherAssistant Agent Expanding LMS integrations to provide native, one-click synchronization with major platforms beyond manual exports (such as direct API integrations with Canvas and Moodle).

Introducing multi-language and localized curriculum adaptation features to support diverse international academic boards and university standards.

Enhancing the AI grading analytics dashboard to provide instructors with macro-level insights into student learning gaps and common misconception trends across semesters.

Built With

• ai
• api
• authentication:
• bearer
• database:
• fallback
• frontend:-react
• integration:
• jwt
• mysql/mariadb
• openai
• pydantic
• react-router
• safe
• sqlalchemy
• tailwind-css-backend:-fastapi
• tokens
• vite
• with
• xampp