CodeE AI - True Peer Programmer

• 

• 

CodeE AI: True Peer Programmer Inspiration

As students preparing for technical interviews, we noticed a troubling pattern in how people were using AI coding tools like GitHub Copilot, Cursor, and ChatGPT.

These tools are incredibly powerful. They can write entire solutions to LeetCode problems in seconds. That power creates a new problem:

You ask ChatGPT for a solution. It gives the full answer. You paste it, submit, get a green checkmark, and move on.

You have “solved” the problem, but you have not actually learned how to solve problems.

These tools are optimized for productivity, not learning. They are brilliant for professional developers who already know how to code and want to move faster. For students learning algorithms and preparing for interviews, they can become counterproductive.

We asked ourselves:

What if AI was designed specifically for learning? Not an AI that writes code for you, but one that teaches you to think like an engineer.

That idea became CodeE AI.

What it does

CodeE AI is a voice-powered coding mentor that uses a different approach to AI assistance.

Instead of giving complete answers, CodeE AI uses Socratic dialogue.

You choose a LeetCode problem and start coding. When you get stuck, you click a button and speak to the AI.

Example:

“Can you help me with this problem?” “What data structure could help you track which elements you have already seen?” “Maybe a hash map?” “Good thinking. What would you store as the key and what would you store as the value?”

The system automatically knows:

which problem you are solving

what code you have written so far

the programming language you are using

The frontend syncs your problem context to the backend every five seconds. When the voice session starts, the AI already has the full picture.

You do not paste code or explain the prompt. You just talk and reason. The AI guides you instead of solving it for you.

How we built it Frontend: React with Monaco Editor

We built a React application using Monaco Editor, which powers VS Code.

Students get:

syntax highlighting

autocomplete

curated LeetCode workspace

A background task sends updates every five seconds containing:

problem ID

problem title

current code

language

These are posted to the backend transparently.

Backend: FastAPI WebSocket Proxy

The backend performs two key roles:

Stores session context

Acts as a proxy WebSocket server

When a student starts a voice session, their browser opens a WebSocket connection to our FastAPI backend. The backend then opens a second WebSocket connection to ElevenLabs Conversational AI.

Audio flows both ways:

Frontend → Backend → ElevenLabs (user speaking) ElevenLabs → Backend → Frontend (AI responding)

Context injection strategy

The hardest part was teaching the AI what problem the student was solving.

Attempts:

metadata on connection → ignored

overriding system prompt → rejected by ElevenLabs

Working solution: We inject the problem context into the first user message.

Example: When the user says:

“Can you help me?”

The backend transforms it into:

“[CONTEXT: I am working on ‘Longest Substring Without Repeating Characters’ and here is my code: …] Can you help me?”

To the AI, it appears as if the user naturally provided context.

Tech Stack

React 19 (TypeScript)

Monaco Editor

Zustand state management

FastAPI backend

WebSockets for streaming audio

ElevenLabs Conversational AI

PCM audio, 16-bit, 16 kHz

Challenges we faced The “Two Sum” bug

No matter which problem users selected, the AI kept thinking the problem was Two Sum.

Root causes:

backend sent problem_id instead of problem_title

ElevenLabs prompt contained a hardcoded Two Sum example

context injection timing was wrong

Fixes:

stored and transmitted problem titles

removed the Two Sum example from the prompt

corrected injection logic

Result: the AI now understands the correct LeetCode problem.

WebSocket library upgrade

Upgrading websockets to version 15.0.1 broke voice streaming with:

TypeError: connect() got an unexpected keyword argument 'extra_headers'

The parameter changed to additional_headers. One-line fix, one hour of debugging.

TypeScript errors

Examples:

NodeJS.Timeout used in browser code where setInterval returns a number

localStorage.setItem receiving possible null values

They were frustrating during development but prevented runtime crashes.

React duplicate keys

We had 81 total problems but only 79 unique problems. Duplicate keys created console warnings. We resolved this with Map-based deduplication.

Accomplishments we are proud of

real-time problem and code synchronization

sub-500 ms conversational latency

tutor-style guidance prompt

clean, production-ready architecture

79 carefully curated interview problems

What we learned Technical lessons

WebSocket proxy management is tricky

context strategy matters as much as content

live backend logs speed up debugging

Product lessons

people speak differently in voice than in chat

type safety prevents production bugs

clear error messages save huge debugging time

What is next for CodeE AI

adaptive difficulty hints

realistic interview simulation mode

code review and reflection mode

multi-language support beyond Python

progress tracking and mastery visualization

Vision

CodeE AI focuses on learning, not shortcuts.

The goal is not to:

solve problems for students

The goal is to:

help students think like engineers

strengthen problem-solving ability

promote deep understanding

Built With

• fastapi
• python
• react