I'm Lokesh. 12 years of building apps from India.
What I've shipped:
• an AI platform serving 4M+ farmers
• features in an app with 30M+ users
• 20+ mobile and web apps for founders
• https://t.co/Wuym7E2yB8, my own SaaS at $4K MRR
Here I share what I learn about React Native, AI agents, App Store reviews and building small SaaS.
Building something? My DMs are open.
@ReaxMedia@alliance@AppStore Congrats on reaching TestFlight. Beta App Review uses the same checks as full review, so a demo account and one line in the notes on how to reach the main feed saves a round trip.
@VadimStrizheus Works well once the model gets a clean, documented events table instead of raw dumps. Otherwise it confidently writes SQL against the wrong columns. A small layer of 10 to 15 named metrics fixed most bad answers for us.
@yuvrxxj@getintus Tip for the 30th: the first external TestFlight build needs Beta App Review, usually under a day but not always. Upload a few days early with internal testers first, and write clear HealthKit purpose strings since review checks those.
@neon_time Grep the built bundle, not the source, for sk_, service_role and secret. And treat anything prefixed NEXT_PUBLIC_ or EXPO_PUBLIC_ as already public, because it is.
@ammarahm_ed React Native driving a GPU-native component kit on desktop is a big unlock. Curious how text input and IME behave, that's usually where non-native desktop UI breaks first.
@PaulSolt What helped us: cap agents at one booted simulator each and add `xcrun simctl shutdown all` to the end of every task in the agent rules. `xcrun simctl delete unavailable` also clears out a surprising amount of old sim junk.
Voice AI lesson from shipping it in production:
Users don't judge the model. They judge the pause.
Under ~800ms from end of speech to first audio feels like a conversation. Past 1.5s they start talking over it.
Stream everything: STT partials, LLM tokens, TTS chunks.
@rsStats_ Nice set. Are these running as Skia shaders on the UI thread? Dissolve effects are usually where low-end Android drops frames first, so a quick perf note per component would help people choose.
@wcandillon@expo Great news. Skia is one of the few libraries that lets a small RN team ship UI that looks custom-built, so funding that keeps it maintained helps everyone building on Expo.
@kacperkapusciak@figma Seeing the same. Writing code stopped being the bottleneck, knowing what good looks like is. The person who can judge the design, the code and the edge cases in one pass ends up shipping the most.
@Baconbrix The loose preferences tip matches what we see building agents. Hard rules make the model stop at the first miss, ranked preferences let it trade off. For anything that spends money, we still keep one confirm tap before it charges the card.
@milrvari@GooglePlay@STIQVerse Congrats on getting through the 14 days. Be specific in the production access form about what testers found and what you changed. Vague answers there are a common reason first apps get sent back for another round.
@jaysquaredlabs Good fix. On iOS 16+ AppTransaction gives you the environment directly, so you can split App Store vs TestFlight without a server. Also worth excluding the review window, App Review can open a build several times per submission.
@TrninhHuy1@expo@wcandillon@swmansion@nativecookdev I'd keep the speed and add a small stagger so the rightmost digit settles last. The number reads as final before the motion ends. Also worth checking Reduce Motion and just fading the value in for those users.
Supabase tip before you ship:
Your anon key ships inside your app. Anyone can pull it out.
Row Level Security is all that stands between that key and your database. Enable it on every table, then query while logged out.
Get rows back? You're not ready to launch.
@DLKFZWilliam2 The 4.3 one hurts most. What usually gets past it is opening the review notes with the one thing your app does that the others don't, and making that the first screenshot too.
@mudassar6969@expo@supabase@wcandillon Server-side score checks are the part most no-login leaderboards skip, then one spoofed request tops every board. Do you also rate limit per device, or has the Supabase function check been enough to keep it clean?
@slavafomin2 Agree. The decisions that hurt later are the silent ones: where state lives, what runs on device vs server, which table owns which data. Writing those down before the first line of code is cheap. Migrating them after launch with real users is not.
@expo "Agent proposes, you sign off" is the right pattern for AI in consumer apps. Users trust the output a lot more when nothing gets saved without a tap, and a bad generation becomes a non-event instead of a support ticket.
@catalinmpit Open source is a big part of why devs trust a tool enough to put it in production. Going closed now trades that trust for protection against clones that can be spun up in a weekend anyway.