• 

  Tako

• 

  Modes

• 

  home dark

• 

  Report

• 

  sensei

TakoSensei

Inspiration

A friend used to quiz me hard before exams. He gave real scores and showed exactly where I lost marks. Most tools just explain concepts. They ignore the fact that students lose marks because they cannot structure answers the way examiners want.

What it does

TakoSensei turns notes into two modes. Exam Mode runs a fixed 6-question paper worth 20 marks with strict point-by-point grading. Sensei Mode runs up to 20 adaptive questions that test both understanding and answer structure.

How I built it

The server is stateless. All session state lives in localStorage. Every turn sends the full state to the API. The backend runs an intent router, a mark-scheme evaluator, and a deterministic move selector before Tako replies. I generate a fixed mark scheme and model answer for each question at the start so grading stays consistent.

Challenges

Early versions let the model give vague scores. I had to force it to check specific mark-scheme points only. Command-word penalties also broke often because the model wanted to be helpful instead of strict. Managing progress, attempts, and anti-copy rules without a database required careful local state handling.

What I learned

Prompts alone are not enough. I had to treat AI as a pipeline of separate agents with strict JSON contracts and deterministic logic on top.

What is next

I plan to add character animations that react to scores, move to stronger models for the hidden grader while keeping chat responsive, and improve rewrite feedback so it targets the exact structural mistakes each student repeats.

Built With

• api
• css
• css3
• framer
• groq
• html5
• llm
• localstorage
• motion
• next.js
• openrouter
• react
• rest
• serverless
• tailwind
• typescript
• vercel