Context Engine: Giving 60 Engineers One Shared Memory
A small AI experiment at YouVersion turned into a system for the whole engineering org. What used to take weeks of meetings took seconds. Shipped in 3 weeks. Won the company hackathon.
When AI was first being adopted at YouVersion, I saw the benefits fast. The first thing I built was small: an LLM-powered tool that read the code changes in a merge request, matched them to the linked JIRA ticket, and posted a plain-language summary of what actually changed back onto that ticket.
Sounds minor. But it replaced something that used to be manual and easy to skip: a developer translating their own technical change into something a non-technical project manager could actually use. Automating it meant that connection stopped depending on anyone remembering to write it down.
A bigger version of the same problem
That small tool pointed at something bigger. YouVersion's engineering org has over 60 engineers across nearly 95 production applications. The domain space is huge, and the knowledge for it is scattered across a handful of people's heads at a time.
Getting up to speed on a domain you didn't own meant the usual path: find the right people, set a meeting, hope it covers enough ground, and if it doesn't, set another one. Multiply that across a big org and questions start taking weeks to answer. Some never get answered, and the project attached to them just dies.
I kept coming back to one question: what if a developer, or a project manager, could just ask an LLM and get an answer grounded in the actual domain, in seconds? That became Context Engine.
What it does
Context Engine explored and indexed codebases across the org, then documented patterns and stored concepts: org-wide authentication conventions, tribal knowledge specific to a project, code practices, tech stacks. Developers and project managers could log in and get answers in seconds.
How it's built
Retrieval. Concepts were stored and indexed as vector embeddings on Postgres. Retrieval ran on cosine similarity against the user's question to pull back what was actually relevant.
Context compaction. Long conversations run into a real limit: tokens. Past a threshold, I sent the whole current context to a cheap model, Haiku, to summarize. The summary got stored in a summaries array, the working context got cleared back to the last few messages, and every turn after that sent the summaries plus the recent messages. That's a compaction system.
Tools. Built on the Vercel AI SDK. Tools for GitLab exploration, web search, concept storage and retrieval, JIRA integration. Tool access was gated per user through the auth system I'd already built for YouVersion's web apps, so what an agent could reach matched what its user was actually allowed to see.
What shipped
Context Engine went from idea to production in three weeks and won the company hackathon. It's the project I point to when someone asks about AI in production. It closed a real gap between the people who had the answer and the people who needed it.