Bending Spoons

How would you build semantic search for 9 billion notes? Evernote’s AI engineers went deep into the retrieval stack: ENML→ Markdown conversion, embedding model evaluation, reranking, filtered kNN search, Elasticsearch vector indices, inference serving, autoscaling, and indexing billions of notes. Some numbers: • 5,700 notes/sec indexing throughput • 300 GPUs during backfill • 1.8 TB of vectors And surprisingly, the biggest bottleneck wasn't AI processing... Read the article to get into the gritty details of the build with AI engineers Ludovico, Andrea, Fabio, and Alberto. Link in comments. 👇 🔗

London heatwave protocol: team sports in the park, Pimm's in hand. ☀️ P.S. Yes, we complained about the heat. Yes, we’re already missing it. 🥲

Managing the Vimeo finance team. Planning upcoming cash flows. Completing an unexpected audit (the good kind 🎂)... It’s a day in the life of Lisa!

🚨 Wanted: Junior talent. 🚨 Students who are hungry for the next challenge, you don't need to wait until graduation to apply. We offer flexible start dates and part-time arrangements, so you can join us without sacrificing your education. If you have strong drive, passion, and problem-solving skills, apply through the link in comments. 👇

600 hours saved every day. 👏 The StreamYard team just shipped a major optimization to AI Clips. Previously, 12 minutes of each generation was spent reframing videos. David cut that by 75%, saving creators 9 minutes per generation and 600 hours collectively every day. For StreamYard, this more than halved compute time on GPUs, translating to over $800k in savings every year. Time saved for users. $$ saved for StreamYard. 💪

Calling business students in Europe... 🔊 This edition of First Ascent is for you! Join us in Milan this October, all expenses paid. More info and apply: link in comments. 🔗

Students and recent grads in Milan, this is for you! 📣 Applications are open for the next Follow the White Rabbit Aperitivo. Join us on June 19 at our Milan HQ for an evening of tech chats, cibo, and cin-cins. Apply now: Link in comments. 🔗

How many engineers does it take to build a feature end-to-end? Just one. Komoot engineer Simone developed Peak Bagging—a tool to track every summit you conquer—solo. Database, backend, Web, Android, and iOS. Our engineers are T-shaped. That means developing skills horizontally and owning goals, not technologies. It’s a bigger upfront time investment. But in the long term, we’ve found it unlocks greater impact. What should Simone ship next?

Brightcove’s biggest redesign in years: Prism. The new UX was created to make complex workflows simpler. Common tasks—publishing, editing, reviewing performance—are easier to access and take fewer steps to complete. Plus, a unified design system runs across every screen, with a refreshed visual style, new typography, and more efficient use of space. Watch the video to see Prism in action. 👇

We made Vimeo 1.7x faster. ⚡ Everyone hates frustrating load times. We’ve been on a mission to improve them. To do this, the team mapped every moment—from opening the home page to uploading a video—and ranked them by the number of users affected. Then fixed them in order. Here are the key changes: 🏡 Home and Manage Video pages Vimeo used to load everything at once, meaning front-and-center videos could be queuing behind invisible background elements. Now, visible content loads first. 📹 Video uploads Previously, video uploads were processed one at a time, so ten short clips could be stuck behind one hour-long video. We rebuilt the flow so videos upload in parallel and appear as soon as they're ready. 🔎 Library search We rebuilt the search pipeline to return results faster by reducing the amount of irrelevant content it combs through. 🏗️ Behind-the-scenes improvements The team also released a bunch of architectural changes and infrastructure upgrades, boosting speed across every page. These are just the early results. There’s more to be done to make Vimeo lightning fast. Onwards! 🚀