CodeSense

Homepage
Results Dashboard
Finding Detail

Inspiration

Every developer wastes 2-4 hours daily hunting bugs that a smarter tool could catch in seconds. Traditional code review tools only flag surface-level issues — they never trace root causes across files or explain why code is wrong. Software bugs cost the global economy $85 billion per year. We built CodeSense to change that.

What it does

CodeSense is a multi-agent AI system powered by Amazon Nova 2 Lite that autonomously reviews entire codebases. It detects bugs, security vulnerabilities (OWASP Top 10), and performance issues — then explains root causes in plain English with ready-to-use code fixes.

Key features:

4 specialist Nova agents running in parallel — Bug Hunter, Security, Performance, Validator
Self-correction loop — Validator Agent reviews and eliminates false positives
ML anomaly detection — IsolationForest computes a composite Bug Risk Score per file
Chat interface — ask Nova follow-up questions about any findings.

How we built it

AI Core: Amazon Nova 2 Lite via Amazon Bedrock — reasoning, tool use, 1M token context
Agents: Custom orchestration with parallel execution using Python ThreadPoolExecutor
Backend: Python + FastAPI on AWS Lambda
Data Science: scikit-learn IsolationForest — Composite Risk Score = 40% anomaly + 30% complexity + 30% agent confidence
Frontend: React + Vite with real-time agent progress, ML heatmap, Nova chat interface
Code Analysis: AST parsing + GitHub API for repo fetching

Challenges we ran into

Python 3.13 compatibility issues with scikit-learn required upgrading to version 1.5.2 to support the latest Python runtime
Amazon Nova 2 Lite free tier token limits were exhausted during development and testing, requiring us to implement a fallback AI provider and optimize our prompts to minimize token usage per agent call
Tuning the self-correction loop to accurately eliminate false positives without removing real bugs required careful prompt engineering
Optimizing Nova API calls to balance reasoning depth with response time — finding the right max_iterations and temperature values for each specialist agent
Implementing parallel agent execution with proper error handling to prevent one agent's failure from crashing the entire analysis pipeline
Managing rate limits across multiple AI API calls when running 4 agents simultaneously required implementing sequential execution with delays

Accomplishments that we're proud of

Built a genuine multi-agent system — not just a chatbot wrapper
The composite Bug Risk Score (ML + complexity + agent confidence) is an original contribution
Self-correction loop significantly reduces false positives compared to single-agent analysis
Successfully demonstrated that ML and LLM approaches are stronger together

What we learned

Amazon Nova 2 Lite's tool use capability is extremely powerful for agentic workflows
Multi-agent architectures with self-correction produce significantly better results
Combining IsolationForest with LLM reasoning creates a more accurate and explainable system