Founder, Reelvyn · video ad creator. Short video ads for brands, apps & agencies: TikTok, Reels, YouTube Shorts. Cut from your own footage. DMs open for quotes.
Hot take: encrypted secrets aren't your safety net. Fast rotation is.
Railway just told all users to rotate every env var after a breach.
If a key is pasted in 5 places, you can't rotate it in an hour.
Could you rotate every key today? Yes or no? #DevOps#Security
Hot take: a currency in a column name is a schema with an expiry date.
Moving a platform's home currency to KES meant renaming 31 money columns, a ledger trigger and the quote APIs.
Integer minor units + a currency code. Or decimals? #Fintech
Hot take: porting the TypeScript compiler isn't the hard part anymore. Staying compatible is.
Cloudflare's tsrs (Rust port): up to 8x faster, ~70% less memory vs TS 7.0. Still experimental.
Fast port in your editor, official tsc in CI?
#TypeScript#Rust
Hot take: Kenya's most common data leak isn't a hack. It's "nionyeshe message."
M-PESA Jumbe fixes it: a receipt you can show, with no name, number or balance.
Dev lesson: pick the reader first, show only what they need.
Per-audience views or one full record? #MPesa#Privacy
Hot take: an admin toggle your token check never reads is a placebo.
"Force password reset" saved a flag. Existing JWTs and refresh tokens still worked.
Fix: block all but change-password, kill refresh tokens, clear auth cache.
Stateless JWTs or check every request? #AppSec
Hot take: AI won't make software safe. It'll make "we'll patch next sprint" fatal.
PHP2Shell: 2 critical PHP RCEs, AI-assisted discovery, exploit on mod_php fires before any app code runs. No full mitigation, only the update.
Patch in hours or wait? #CyberSecurity#PHP
Hot take: the best AI support agent knows when to quit.
Mogo Kenya says a phone menu + AI voice agent resolve up to 80% of routine calls (its own figure).
The real win: a missing M-Pesa payment goes to a human, with a summary.
Always offer a human? Yes or no?
#AIAgents
Hot take: "save to DB, then push to the queue" is two systems and zero transactions.
Redis blipped, the order committed, its follow-up job never got queued. Silent.
Fix I shipped: outbox row in the same transaction + 10s sweeper.
Outbox: essential or overkill? #SystemDesign
OpenAI kept the name. DeepSeek is doing the open part.
DeepInfer, the production inference stack behind DeepSeek-V4.1-Flash, is now open source for anyone to run.
Which lab is really the "open AI" in 2026?
🚀 Introducing DeepInfer, our in-house inference framework, now fully open source!
Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.
🔗 https://t.co/JGAik1kgk0
1/4
Unpopular opinion: your next outage will start as a fix.
Google's report on 18h of Cloud Storage errors: an "optimization" bypassed capacity-aware routing, then an incomplete rollback brought it back the next day.
Config changes: canary like code, or fast-path? #SRE#DevOps
DeepSeek just open-sourced DeepInfer, the stack serving DeepSeek-V4.1-Flash in production. Not a demo, their real engine.
Lower compute costs, less energy, and anyone can learn how inference runs at scale. Huge for builders.
Who's trying it first?
🚀 Introducing DeepInfer, our in-house inference framework, now fully open source!
Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.
🔗 https://t.co/JGAik1kgk0
1/4