More merge requests doesn’t mean more value. With 120+ devs shipping faster with AI, we couldn’t tell new features from fixes to our own bugs. So we had an LLM sort every MR into five categories, every day. builders.fullscript.com/post…

When should you ask AI, and when should you ask code? We tried an LLM, a custom model, and pure OpenCV to crop Supplement Facts from thousands of labels. Each failed alone. Together they saved 88+ hours. builders.fullscript.com/post…

Our search tuning loop used to be: fix “magnesium,” break three other queries, repeat. The fix wasn’t a better boost. It was turning views, carts and purchases into training data for a ranking model. Written entirely in Ruby. builders.fullscript.com/post…

It started as a way to pull data from a few places into one. A few long sessions with a Staff Engineer, a limited beta, and several security reviews later, it became EKG: one door between our AI tools and everything else. builders.fullscript.com/post…

we loved the showcase room @railsworld and having the opportunity to chat with vendors like @PlanetScale

proud of Markle for speaking @railsworld rails at scale.

if you see any of us at the Thirstaid station @railsworld , hydrate and say hi. repping @FullscriptHQ

Sending off @patgvice @jfouchard and Andrew to enjoy Rails at Scale @rails

We are so happy to add another little one to the Fullscript Eng team.

what does it take to know if a 2am security alert is real? for us it was 30 to 60 min of manual digging, per alert, across hundreds a month. we built a pipeline that now does that first pass in under 5 minutes. builders.fullscript.com/post…

a page reload used to sometimes show a stale conversation. rare, until new features were about to make it common. so we gave every conversation one owner in postgres, and the next message picks up from the last save. builders.fullscript.com/post…

What happens to "human in the loop" when attackers exploit new flaws in hours, not months? An approval queue becomes a slower way to lose. Here's where we put human judgment instead: builders.fullscript.com/post…

We benchmarked AI coding agents against our own git history; real commits, real review. All three Claude models landed within 4.5 points of each other. Close enough that cost, not capability, should drive which one you pick for a given task. Read more: builders.fullscript.com/post…

We added controlled experiments to our Flipper feature flags without touching a single call site. Two mechanisms: a second independently salted hash, and exposure tracking that waits until someone actually reaches the feature. 70 rollouts later, here's how it works. builders.fullscript.com/post…

Models degrade well before they hit the context limit. Attention thins, instructions at the edges get lost. So we stopped asking "does this fit?" and started asking "is this worth the space it takes on every future turn? Part 2 of Building Fullscript Assist is out: builders.fullscript.com/post…

Making maintenance tasks easy to run meant we ran a lot of them. That surfaced silent failures, vanishing logs, and dead code piling up. Part 3: the observability and self-cleaning governance we built around Shopify's maintenance_tasks gem. builders.fullscript.com/post…

Can't wait until next year!

Our IT team's first Builder's Corner post: two warehouse network builds, four WAN paths, and a design philosophy borrowed from hospitals. Redundancy pays for itself the first time something fails and nobody notices. builders.fullscript.com/post…

First day @rubyconf and proud to be a platinum sponsor and give a talk on Teaching Ruby to Rank. @FullscriptHQ