Replying to @Kalshi
@Kalshi's live chat grew 14x across the NBA Finals, the World Cup, and NFL opening weekend, powered by Stream Chat and AI Moderation.
59,000+ concurrent connections at the NFL peak.
18 ms average moderation check.
1 engineer on the chat build.
gstrm.io/4A1yRkC
Nov 7, Nuremberg. Our @stefanjblos talks local AI and what you can already do with it at Google DevFest. 10:00 to 19:00 CET, Zollhof 7.
gstrm.io/4yMjE5X
3
809
Replying to @Rosterfy2
@Rosterfy2 built volunteer chat on Stream.
42,000 volunteers connected at a major global event.
80,000 messages across 795 channels.
10,000 to almost 40,000 monthly active users, April to May.
gstrm.io/47HTPZT
2
4
435
Our guide to moving off Perspective API, written by GTM Lead @kenziewilsn.
65 days until it shuts down.
5 steps to move to Stream.
gstrm.io/4BC0wuf
Our Software Engineer @adityaalifn built an AI agent that investigates production issues for our chat team. Every claim cites a query or a line of code, and only engineers can approve a fix.
Inside the numbers: gstrm.io/3W3cyMO
4
17,702
6 features every social trading app has converged on, and how each one is built: verified profiles, a live trade feed, trading from the feed, chat, moderation, and live video.
gstrm.io/3W3InVJ
Our fintech build guide:
2 Postgres triggers: append-only, always balanced.
0 card numbers on your servers.
1 database transaction per transfer.
Read more: gstrm.io/4eGXquR
4
2,542
268,681 chat events a second delivered at the peak of an NFL Sunday, with zero 5xx errors on the tracked endpoints. 16.7 messages sent a second, each delivered 2,355 times.
1,500+ text moderation requests per second at peak, with API latency under 50 ms. Measured on seven days of production traffic.
Didn’t have weekend plans? Now you do. Amazing build by @FloorRene
Meet Deep Signal 📡
A race game for game night: the board on your TV, a controller on every phone. Everyone programs 5 moves in secret, then every drone runs at once.
Built with @ServerpodDev and @FlutterDev for the Serverpod hackathon. A thread 🧵
6
1,034
15 ms median feed reads with 4.5 million users online at once. Measured in our benchmark test.
For trading apps, that means the alert, the discussion, and the live briefing reach users while they still matter.
See how Stream's architecture handles it: gstrm.io/4rWafak
2
203
iPhone Duo support for Stream Chat SwiftUI, tested and confirmed by @amos_gyamfi, with a sample app to try it yourself.
I just tested iPhone Duo support for StreamChat SwiftUI. It looks great.
Your iOS messaging app with StreamChat SwiftUI v5.12.0 is now ready for iPhone Duo.
The channel list, messages UI, and navigation automatically adapt to iPhone Duo’s display sizes, orientations, poses, and reserved regions. Want to test it quickly? 👇
Same flag count, opposite fix. Ban the 1%. Nudge the rest.
Replying to @kenziewilsn
The first kind of problem is one person. Same account, every night, in the chat being awful. On the live shopping marketplace Stream moderates, 1% of accounts set off 17.6% of every moderation action.
The second kind is a big group of people. The biggest thing we flag there is "DM me on Instagram." 6,115 messages in ten days, from thousands of different people and three quarters of them only did it ONCE.
2
476
Our DevRel Max is hitting the road again, heading to AI Builders Berlin on September 22nd.
Tickets are sold out, but get on the wait list to see if a spot might open up: luma.com/ber-sep22
Fourteen games, one Sunday, 38,900 moderation decisions. The number that matters is the last one.
Fourteen NFL games yesterday. One prediction market.
38,900 items into the moderation queue. 32,563 actioned automatically. 1,600 looked at by a person.
Zero false positives.
That ratio is the only reason any of this works. You cannot hire your way to 38,900 decisions in a day, and you cannot get them wrong either.
2
526
Our engineering teams are hard at work 👀
9
1,545
Great to see Pratim take the stage at RTC.ON 2026 this week. His talk, "SVC on Android, Debugged," traces a real debugging journey: Android's hardware VP9 encoders accept SVC configuration but silently produce only one spatial layer instead of the promised three, creating a mismatch between what the SFU expects and what actually shows up in the stream.
Take a look if you work in video on Android: youtu.be/YBOGK3iwzdA?t=10757
4
25,407




















