Fast Code Generation Is Easy. Safe System-level Change Is Not.

Hey, we'll discuss scaling fast and engineering lessons learned in the last 15 years of tech startups. Scaling involves three things: business, team, and tech. Business scalability relies on sales and customer acquisition costs. Engineering is a tool the business uses. Scaling the team is vital as tech problems are often people problems. Team structure affects architecture and product development process. Organize teams based on purpose, not technology. Spend less time being blocked by other teams. Ship features without getting blocked. Own your own mess. Focus on product engineering partnership. Build faster using feedback cycles. Build appropriate solutions for your use case. Let go of ego and experiment with different approaches. Engineers own their own mess. Avoid work in progress. Finish the work and focus on fixing it later. Have a conversation before writing code. Scaling the tech is easier than you think. Pick an off the shelf design. Save innovation for core parts. Pick existing solutions. Focus on solving the problem. Don't waste time trying to predict future scale. Scale will surprise you. Do what works for your business. Push back on unnecessary complexity. Understand the cost of ideas. Modify the situation to fit existing design. Architecture is like a dependency graph on your code. Reduce architectural complexity by organizing code based on what it does. Use vertical models and avoid creating excessive dependencies. On the client, use vertical modules. On the back end, consider a service-oriented architecture. Start with a monolith and transition to microservices if necessary. Use folders instead of microservices when you have a small team. Use vertical models and contract or type-driven development to define clear APIs and interfaces. Avoid highly interconnected code and duplication. Focus on data structures to avoid complexity and the need for translation layers. Building translation layers can lead to slow user experience. Vertical teams aligned with vertical code allow for fast problem-solving, full control of features, and efficient data handling. Understanding the entire domain enables faster development with fewer bugs.

This Talk discusses automating code changes for Android repositories, utilizing tools like JSCodeShift and Abstract Syntax Tree. The speaker shares a real use case example of maintaining a design system library and making changes to a component. The talk emphasizes the importance of automating repetitive tasks and using the power of abstract syntax tree for code changes. The Q&A session covers topics like source code formatting, TypeScript support, and cultural embedding of code mods. The talk concludes with insights on when automation is worth it and the limitations of code mods for monorepo changes.

Today's Talk focused on code architecture, modularization, and scaling in software development. The speaker discussed the benefits of separating code by domain and using tools like NX to improve productivity and enforce modular architecture. They also highlighted the importance of automating library creation and configuration. Additionally, the Talk covered code scaling and deployment strategies, including caching and automated code migrations. The speaker emphasized the flexibility and scalability of Fastify and the advantages of using a monorepo for front-end and back-end development.

Cloud Code introduced sub-agent feature for better performance. Understanding the sub-agent concept and its role. Challenges of sub-agents in direct implementation. Parent agents' limited visibility. Efficient context management with MD files for sub-agents' tasks and reports boosting Cloud Code performance. Guidelines for creating useful sub-agents with specific MCP tools. Utilizing TweakScene MCP for scene design reference and defining rules for sub-agent tasks. Setting goals, output format, and rules for sub-agents. Building a chat GPT replica with chat.cn UI and Versatile AIS. Integration with VerCell AI SDK for seamless implementation.

This Talk discusses the journey from thousands of Lint failures to zero in a codebase at Linton that is over 80 years old. The approach involved implementing rules, incentives, and tooling to address the issue. The tool called Checkup was used to visualize ESLint failures by team and lint rule, providing accountability and responsibility. The efforts resulted in cleaning up over 6,000 lint failures, with 55 contributors, and a 30% increase in perceived code quality.