Inspiration
Every computer science student and self-taught programmer encounters a moment of complete frustration when facing cryptic compiler errors, segmentation faults, or silent runtime crashes. While modern AI code assistants can output fixed code, they often rewrite the entire script without explaining why the original code failed. This creates dependency rather than understanding. We built CodeFix AI to shift the paradigm from mere code generation to genuine code comprehension—empowering learners to understand root causes, master core concepts, and debug independently.
What it does
CodeFix AI is an intelligent developer assistant and pedagogical debugging tutor powered by NVIDIA Nemotron on Nebius Token Factory:
- Multi-Language Diagnosis: Pinpoints bugs across Python, C++, JavaScript, Java, C, and HTML/CSS.
- Root Cause & Minimal Fix: Identifies the exact failure trigger while preserving the user's original logic and structure.
- Side-by-Side Comparison: Visually contrasts original buggy code with corrected code, complete with single-click copying.
- Adaptive 3-Tier Explanations: Tailors explanations to the student's mastery level:
- Beginner: Simple analogies and zero-jargon intuition.
- Intermediate: Data structure boundaries and control flow mechanics.
- Advanced: Runtime memory semantics, language specifications, and complexity.
- Static Test Case Generator: Synthesizes standard, edge, and boundary test cases to validate logic safely.
- "Teach Me This Error" Deep-Dive: Teaches the underlying computer science concept (e.g., fencepost bugs, pointer safety, asynchronous race conditions) with reusable code patterns.
- Zero-Retention Privacy: Operates with ephemeral session memory without permanent server storage of student code.
How we built it
- AI & Cloud Inference: Integrated the OpenAI-compatible Nebius Token Factory endpoint (
https://api.tokenfactory.nebius.com/v1/) with high-throughput NVIDIA Nemotron models, usingnvidia/llama-3.1-nemotron-70b-instructas our primary model for superior instruction-following and reasoning. - Backend Architecture: Modular Python 3 backend structured with robust validation, safety checks against destructive commands, and a multi-stage fallback JSON parser.
- Frontend & UX: Developed in Streamlit with custom dark charcoal CSS (
#0d1117), structured metrics cards, and intuitive side-by-side diff views. - Quality Assurance: Built and verified an automated 13-test Pytest suite covering input bounds, schema conformance, client diagnostics, and fallback logic.
Challenges we faced
- Strict JSON Schema Compliance: Ensuring the model strictly outputs clean JSON without conversational preamble required designing a disciplined system prompt with strict output boundaries and building a resilient fallback parser.
- Pedagogical Balance: Prompting the model to fix code surgically without completely rewriting the student's variable names or style took iterative prompt engineering.
- Execution Safety vs. Test Generation: Balancing the desire to test code with student safety—we implemented static reasoning test case synthesis to protect server environments from arbitrary code execution.
Accomplishments that we're proud of
- Successfully routing real-time inference through Nebius Token Factory with sub-second response times.
- Building a truly useful pedagogical assistant that adapts to 3 distinct skill tiers rather than treating every learner the same.
- Achieving 100% test pass rate across 13 comprehensive unit tests.
What we learned
- How to harness NVIDIA Nemotron’s RLHF-tuned reasoning capabilities for precise structured code analysis.
- How to implement high-throughput, low-latency LLM inference using Nebius Token Factory’s OpenAI-compatible APIs.
- The importance of defensive parsing and graceful degradation in production AI interfaces.
What's next for CodeFix AI
- In-Browser Sandboxed Execution: Integrating WebAssembly (Pyodide / WebContainers) to safely execute generated test cases directly on the client machine.
- IDE Extensions: Bringing CodeFix AI diagnostics into VS Code and JetBrains IDEs.
- Multi-File Workspace Analysis: Supporting repository-wide import analysis and Git diff debugging.
