How to Improve AI-Generated Code with SonarSweep

on YouTube ↗13m

" summary

SonarSweep uses Sonar static analysis to analyze, deduplicate and automatically remediate training code, then fine‑tunes models on a curated corpus of 70K Java samples. Fine‑tuning yields concrete gains: a Llama variant shows 67% fewer vulnerabilities and 32% fewer bugs on benchmarks, a closed‑model run sees ~52% fewer vulnerabilities, and overall experiments report ~41% reductions in bugs and vulnerabilities and an 80% drop in code smells; demo model is on Hugging Face.

AI-written summary · about

" chapters (23)

1. 0:00Introduction: Improving AI Code Generation Quality
2. 0:17The Problem SonarSweep Solves in AI Code Generation
3. 0:35Enterprise AI Adoption and the Growth of Technical Debt
4. 1:16Why LLMs Generate Bugs and Security Vulnerabilities
5. 1:40Scaling Models Isn’t Enough: Fix the Training Data
6. 2:19Garbage In, Garbage Out in AI Code Generation
7. 3:06Poisoned Training Data and Hidden Backdoors in Code
8. 3:42SonarSweep: Treating Training Data Like Production Software
9. 4:02Using Static Analysis to Curate Secure Code Datasets
10. 4:25SonarSweep Pipeline: Analyze, Deduplicate, Detect Issues
11. 4:48Automated Code Remediation to Fix Bugs and Vulnerabilities
12. 5:09Filtering Low-Quality Code and Increasing Dataset Diversity
13. 5:40Benchmarking AI Models for Code Quality and Security
14. 6:32Measuring Code Quality: Security, Reliability, Maintainability
15. 7:20Fine-Tuning an LLM with 70K High-Quality Java Samples
16. 7:41Results: Fewer Bugs and Security Vulnerabilities in Generated Code
17. 8:19Maintaining Functional Correctness After Fine-Tuning
18. 8:51Releasing the Demo Model for Hands-On Testing
19. 9:37Example: Fixing Database Transaction Memory Cleanup
20. 10:43Demo Metrics: Bugs, Vulnerabilities, and Code Smells
21. 11:15No Overfitting: Preserving General Programming Ability
22. 12:08Key Takeaway: Training Data Quality Is a Critical Asset
23. 12:52How to Try Sonar Sweep: Blog, Model, and Early Access

" related talks

• 19mWith great power comes great responsibility in code qualityManish Kapur · Microsoft Build 2024
• 49mClean Code in the AI eraJonathan Vila Lopez · JDD 2024
• 46mCode Quality in the AI era, needed more than everJonathan Vila · Voxxed Days Bucharest 2025
• 15mUsing AI effectively in large Java codebasesNupur Agarwal · JCON EUROPE 2026
• 47mSeven Habits of Highly Effective AI Java CodingJonathan Vila · JCON EUROPE 2026
• 33mAccelerate Code Quality in the AI AgeAdam Tornhill · Øredev 2024
• 33mShipping Fast, Breaking Silently: Why AI-Generated Code Needs a Second OpinionStephen Gray · Vibe Coder Conference 2026
• 09mHow Much Trust Do You Have with LLM-Based Solutions?Matthew Salmon · GOTO Amsterdam 2024