AI Dev Podcast #3 / Как сделать Code Review инструмент / Опыт Т-Банка

How to grow a company-wide product from an AI prototype for geeks? Real-life experience at T-Bank / AI Dev Conf Podcast #3 AI speeds up development, but at the same time, the number of bugs triples, and code generated by neural networks requires double-checking. How can we stop playing with prototypes and build a production solution that will speed up and secure the process? How can we prove its value in terms of time and money when teams resist change? In this episode, we explored a real-life case study at T-Bank: the journey of the AI ​​Code Review tool from an R&D experiment and whitepaper review to a product deployed to thousands of developers. Timecodes in the description 📌 Guests: Nadezhda Egoshina (Product Head, T-Bank). Developing a product-based approach to developer tools. — Georgy Mkrtchyan (Head of the Engineering Software Development Team). His team developed the AI ​​reviewer prototype and handed it over for development. — Andrey Dmitriev, Partner at JUG Ru Group. — Andrey Burakov, Author of the Yet Another Analyst channel, @another_sa. What they discussed (timecodes): 00:00:00 About code review automation 00:00:54 AI accelerates development and changes processes at various levels. 00:01:46 Guest Introductions 00:03:40 Development Trends 00:06:08 How AI Review Works 00:07:56 Integration with Systems 00:10:56 Project Scope 00:11:54 How AI Code Reviewer Began 00:12:47 Market Research and Prototyping 00:14:23 Expectations vs. Reality 00:15:21 The Role of Human Code and AI 00:17:42 Problems with Bugs and Vulnerabilities 00:18:28 Product Development and Metrics 00:20:01 Trust and Expectations 00:23:21 Creating a Framework for Offline Assessment 00:24:18 Hypothesis Testing and Feedback 00:26:14 Transition to Online Assessment 00:31:21 Key Metrics 00:33:39 Comment Analysis and Their Lifecycle 00:40:10 Automatic Fixes 00:42:55 Metrics Analysis 00:46:19 The Risks of Over-Reliance on Tools 00:49:07 Impact on the Merge Request Lifecycle 00:50:47 The Role of Metrics and Trust in Tools 00:52:30 AI Agent Ecosystem 00:54:27 Model Training and Synergy 00:56:37 External Code Review Solutions 01:00:09 Resistance to Change 01:05:28 Monitoring the Market and Competitors 01:07:15 Vision of the Future Product 01:14:30 Summary and Farewell Key thoughts of the episode: 1. You can't just "roll an AI bot" and embed it in GitLab. We'll have to restructure the code review process, teach teams to trust (but not blindly), and change the development culture. 2. The hardest part of product work isn't metrics and architecture, but dealing with people's resistance. You'll have to constantly prove its value and gather feedback through surveys and ambassadors. What you can learn today: — A recipe for turning a prototype into a product and handing it over to the product team (metrics, offline benchmarks). — Understanding whether to buy a ready-made solution (GitHub Copilot) or build your own, based on security requirements. — Practical advice on test automation and unit test generation. A wish from the guests: Be open to change. AI has already changed the rules of the game, and the worst thing that can happen to a team is to stick with the old processes, thinking that "everything works for us anyway." Related links: — Veracode: AI-Generated Code Security Report. https://www.veracode.com/blog/genai-c... — Aikido vs. Coderabbit Comparison. https://www.aikido.dev/comparison/aik... — JPoint Implementation Report. https://jpoint.ru/talks/20009812-ai-i...