Article URL: https://home.robusta.dev/blog/you-really-shouldnt-copy-paste-errors-into-claude-code Comments URL: https://news.ycombinator.com/item?id=48725359 Points: 7 # Comments: 0
This is an AI-generated summary. ShortSingh links to the original source for the complete article.
Ryan, CTO at airCloset, has detailed how he solved the semantic search problem in a knowledge graph built from 46 production code repositories. Drawing on a prior internal project called db-graph — which made over 1,100 database tables searchable in natural language using AI-generated embeddings — he applied the same pattern to the code graph. By joining the two graphs, the code graph automatically inherited database context without requiring new annotations. For remaining boundaries such as APIs and UI pages, he opted to add targeted annotations only at boundary nodes rather than across all functions, making the approach feasible for large, established codebases. This selective annotation strategy allows AI agents to retrieve verified boundary-level intent, replacing unreliable inference with structured, searchable facts.
A development team built StyleSense, an AI-powered virtual try-on application, during a 72-hour hackathon to address the problem of online clothing returns, which cost the retail industry over $816 billion annually. The app allows users to upload a selfie and preview clothing items from any product URL on their own image using AI models from Runway ML and Anthropic Claude. The backend runs on FastAPI with Amazon Aurora PostgreSQL Serverless v2 as the primary database, chosen for its ability to scale to zero during idle periods and support IAM-based authentication that eliminates the need to store database passwords in code. The Next.js frontend was deployed on Vercel, with both the frontend and database hosted in the Mumbai region to keep database round-trip latency under 20 milliseconds. Supabase was used separately to handle authentication, real-time features, and social functionality, while Aurora managed core transactional data requiring SQL joins.
A developer built BrewCore, a full-stack espresso tracking application, as an entry for the H0 Hackathon. The app allows home baristas to log shot parameters such as grind size, dose, yield, and flavor profiles, replacing informal methods like sticky notes. BrewCore is built on Next.js 15, hosted on Vercel, and uses Amazon DynamoDB across eight purpose-built tables for efficient data retrieval. An AI coaching feature leverages Amazon Bedrock's nova-micro model to analyze a user's personal brewing history and offer tailored dial-in advice. To speed up development, the creator used a local localStorage mock that mirrored the DynamoDB data structure, enabling offline UI development before switching to live AWS infrastructure.
Custodian Labs has released an open SDK that allows developers to build and deploy AI agents powered by GPT-4o using just a few lines of Python code. The library, called custodian-labs, abstracts away infrastructure concerns such as session management, frontend setup, authentication, and vector database configuration. Once deployed, each agent returns a live, shareable chat URL that can be distributed via tools like Slack for immediate use by non-technical stakeholders. The SDK also supports file-based data sources including CSVs and PDFs, handling chunking, embedding, and retrieval automatically at chat time. Custodian Labs says future updates will introduce multi-agent topic routing and a PII-masking layer aimed at legal and healthcare use cases.
Discussion (0)
Log in to join the discussion and vote.
Log in