This talk was recorded at NDC Copenhagen in Copenhagen, Denmark. #ndccopenhagen #ndcconferences #developer #softwaredeveloper    

Attend the next NDC conference near you: 
https://ndcconferences.com
https://ndccopenhagen.com/

Subscribe to our YouTube channel and learn every day:   
/        @NDC 

Follow our Social Media!

https://www.facebook.com/ndcconferences
https://twitter.com/NDC_Conferences
https://www.instagram.com/ndc_conferences/

#ai #architecure #serverless #genai #llm 

How hard can it be to build an AI-powered code reviewer using modern LLMs? Easier than you think—until it isn’t.

In late 2024, I helped a client launch an AI-powered code reviewer. We built the initial prototype in the first few hours. But we soon encountered many unexpected challenges as we tried to productionize the idea into a coherent user experience.

In this session, let me tell you the surprises, challenges, and lessons learned from building an AI-powered code reviewer and what it tells us about the current state of LLM technologies.

NDC Conferences

Building an AI-powered code reviewer revealed practical challenges beyond the AI itself. The team built a GitHub-integrated tool using AWS serverless (Lambda, EventBridge, DynamoDB) and Amazon Bedrock in 80 hours. Key lessons: published context window sizes don't reflect actual reasoning capability (most models fail beyond 32KB despite claiming millions of tokens), API rate limits and latency create architectural constraints, hallucinations require defensive coding, and unit economics remain challenging for B2C AI products. Cost control required analyzing only changed lines, limiting free-tier usage, and implementing idempotency checkpoints. The AI component was only ~10% of the work—most effort went into handling limits, retries, webhooks, and user experience. Model performance varies significantly by programming language, with JavaScript/Python far better supported than Java/.NET.

How we built an AI-powered code reviewer in 80 hours - Yan Cui - NDC Copenhagen 2025