What Was Filling Up MemoryDB? — Part 2
Hello, I’m Simon (Sangwon Lee), a server engineer on the allocation systems team. This is part two of “What Was Filling Up MemoryDB?” Part one followed our investigation into what…
Buzzvil Engineering
Buzzvil Tech
What we build, what we learn, and how we work.
Explore topics
Find a starting point across engineering, product and AI.
Go embed: Reducing Unnecessary Memory Use with Minimal Changes
Comparing mmap and direct embedding to remove the GeoIP database heap copy found with Pyroscope, reducing unnecessary memory use without changing deployment.
Designers Open the PRs Now
6 months. 432 pull requests. One design team. What got better, and what it cost us. Recently, one of our designers was exploring a lucky-draw concept in the ad participation flow.…
What Filled Our MemoryDB? Part 1: Client-Side Caching and Tracking Items
An investigation that began with an 85% memory alert, traced 57 million unidentified objects, and ultimately found the real cause: ghost keys.
Redesigning Business Systems with AI: Building a CRM to Replace Salesforce
How we built BuzzForce, our in-house replacement for Salesforce: understanding workflows, redesigning the data model, dividing work with AI agents, and moving real business operations into the new system.
From 245 Peak-Time Pods to Around 100: Halving Go Server CPU with DynamoDB SDK v2
Finding a reflection bottleneck with Pyroscope and moving to guregu/dynamo v2: 59% less CPU per request and roughly half as many HPA pods.
A Design System for an Agent
As a Frenchman, I love spending time in the kitchen, cooking things for the fam. As a designer, I love crafting delightful interactions. So in the first half of 2026, I spent a cra…
An Operating Manual Written by the Agent: Five Layers of Harness for AI Coding Agents
Editor’s note. The narrator of this article is not a person. It is Claude Code itself, the AI coding agent that works in Buzzvil’s web monorepo every day. Maxence, who regularly wo…
Improving Feature Flag API p99 Latency—and Contributing the Fix Upstream
Tracing p99 spikes caused by synchronous go-feature-flag flushes holding a mutex during S3 uploads and jitter sleeps, then fixing them with AsyncExporter and an upstream PR.
The AI-Native Design Playbook
Over the past month, Buzzvil’s design team shipped five web applications and made 344 commits. Remarkably, Figma was barely involved. There were no engineering sprint allocations,…
Nobody Owned the Website. Now Everybody Does.
Who owns your company website? It sounds like a simple question. But at most companies, the honest answer is somewhere between "marketing, sort of" and "nobody, really." The people…
The Flip
There's this assumption floating around that AI is going to generate design systems for us. That one day you'll prompt something, and out comes your tokens, your components, your i…
Meet the people behind the work. Browse contributors →