About this session

Documentation is where good intentions go to die. Teams write it once, then it decays silently as code evolves. What if documentation could maintain itself? Quill is an event-driven platform that uses LLMs to automatically generate, update, validate, and summarize software documentation, triggered by the same git events teams already produce. When a PR merges, Quill reads the code changes, identifies affected docs, and opens PRs with precise updates. When a release ships, it synthesizes commit history and deployment data into stakeholder-ready release notes. When docs violate quality standards, it auto-fixes what it can and flags what it can't. This talk dives into the AI engineering decisions that made Quill work in production at scale. We cover how we designed mode-aware prompts using the Diataxis documentation framework to give the LLM a structured target instead of open-ended prose. We explore the no change narration principle, why AI-updated docs must never say previously or was changed to, and how a single prompting constraint transformed output quality. We walk through the multi-pass generation pipeline: per-file LLM calls, aggregation into a cohesive corpus, and information architecture restructuring, and how we handle repos with hundreds of source files without blowing up context windows or costs. Beyond prompting, we cover the system design: an event gateway that fans out git webhooks and package lifecycle events to specialized microservices, a Neo4j knowledge graph that models relationships between repositories, packages, and deployments, and a per-repository configuration system that lets teams opt into exactly the AI actions they trust. Attendees will leave with concrete patterns for building reliable, controllable AI-powered developer tools, not demos, but production systems.

Topics

Speaker

Arjun Harbhajanka

Bloomberg