[2/3] Integrating AI into the development processes of a large company: why allowing all Cursor is not an option (Category Management)
We stopped at past post on platforms and their complexity. What can we do to combat this complexity?
To work on the complexity of individual tools, the guys at Google went to the product side and decided to understand what goals engineers have in software development. They described their approach in the whitepaper “Measuring developer goals”. told earlier. The point is to highlight the phases of the software development life cycle, and within them to determine ala jobs to be done scenarios that engineers need. Further, these scenarios can be digitized using telemetry from platform tools and surveys of engineers. The scenarios themselves are decomposed on the use of several tools and then you can optimize and AI-fix not individual tools, but the entire scenario. (In the future, you can try to give AI agents.) At the same time, it is interesting to look at the expectations of the AI engineers themselves within the platforms. Guys from Google in the summer 2024 Whitepaper "What Do Developers Want From AI?" (review)Three Levels of AI Improvement for Engineers (Interestingly, they drew parallels with the auto industry and drivers.) 1. Enhancing human capabilities Power steering and anti-locking system in cars, autocomplement code in IDE 2. Extending human capabilities Rear-view camera and blind spot control in cars, chat and code review suggests in development 3. Delegating human capabilities - improved cruise control and lane keeping in cars (including Tesla's autopilot)Agency workflow in AI, including automatic deletion of dead code, and test generation In general, this approach shows how AI can gradually penetrate into development processes. And then it is interesting to look at the approaches of Western bigtech companies to the implementation of the second and third levels.
- Google and API design. The Google guys were 2 research papers in the middle 2024 "API Governance at Scale" (my analysis) AI-Enhanced API Design: A New Paradigm in Usability and Efficiency (my analysis). They talked about their approach to API design, which is perfectly standardized. (more aip.dev)And then they tried on top of these standards to screw the api architect. (This is an LLM model for generating speck by product requirements n). - ByteDance and code review. The guys from ByteDance have a scientific article 2025 Bitsai-CR: Automated Code Review via LLM in Practice (my analysis: 1, 2 and 3). In it, they talk about their pipeline to work with code review. 200+ categories, model for generating comments for review, model for filtering generated, as well as feedback loop from engineers for tuning efficiency all systems Uber and large-scale migration from Java to Kotlin. “This Year in Uber’s AI-Driven Developer Productivity Revolution” 2024 (my analysis: 1 and 2). In short, the guys had a lot of Java code that they wanted to rewrite into Kotlin. In a decentralized version, teams would do it sluggishly. 10+ years, the guys screwed first LLM and reduced the score in 3 Once, and then combined LLM and AST and were able to meet the 18 months for the entire codebase migration. Booking and external partnership with Sourcegraph. Booking decided to join AI through an external partner (detail here). Sourcegraph made a QraphQL request for Uber from the Booking API for the text description of the request, made a code review, and also removed a codebase migration project from suspended animation. 14 years. Datadog and AI oncall engineer. Datadog is an observability platform that decided to make an oncall engineer. This is what I told you earlier this year, and I wrote earlier
In final post We'll discuss how things are in T, and we'll conclude with conclusions.
#AI #PlatformEngineering #Engineering #Software #Processes #Productivity