Article Series · Artificial Intelligence
AI in a Closed Loop
AI-Assisted Software Development in Regulated Institutions
An eight-part series on how AI-assisted software development is set up in banks and regulated institutions: the agent era, where local and external models fit, cost and token management, codebase understanding, governance and audit.
- 01
The Problem Is No Longer Copy-PasteCoding agents are not chat boxes; they read the repo, run commands and open pull requests. In a regulated institution the question is no longer "will code leave the building" but "what is an external model authorised to do inside my network". Why control belongs in the system, not in the agent. - 02
Four LayersFrom the outside an AI setup looks like a model and a plugin. In every environment I have built, four layers emerged: model, developer interface, context and system understanding. The options at each layer, what I chose, what I deliberately did not, and a starter setup for a team beginning from scratch. - 03
Tools and the Code GraphTool comparisons focus on "which one writes better code"; in a regulated institution the questions of security, compliance and audit must be answered too. A look at the tools through those questions, eleven questions to ask the vendor, and the job no tool does, the code graph. - 04
Local Model or External Model?How far behind the cloud models does a local model running on our own servers fall? The answer we found by using both: the three reasons for the gap (model quality, context and prompts, trust) and the three questions that decide in a minute which work goes to which model.
The remaining 4 parts will be added here as they are published.