Lech Kalinowski

Ph.D. in Physics ⚛️ | AI & ML Expert. Leading innovations. InCredibles & Innowator Śląska 2024 winner. Bridging deep science & business impact.

[Lech Kalinowski @LeSiOO]

Why it works: Apex is Callstack's LLM and React Native is its home turf. Pair that domain brain with Hermes' autonomy and you get an engineer, not an autocomplete.

Apex: https://t.co/wfLdJDdn7Y
Hermes: @NousResearch

If you ship RN, watch this space. 👀

[Lech Kalinowski @LeSiOO]

Give it one sentence + a deadline. It plans, scaffolds, installs native deps, runs the app on a simulator, screenshots itself, iterates on the visual bar, writes tests, and ships.

Brain: @Callstack's Apex LLM (it speaks React Native).
Hands: Nous Research's Hermes agent.

Unattended, in ~15–30 min each, it shipped a Skia semantic-graph note app with on-device TF-IDF linking, a shader glass-slab toy, and a breathing-orb guide — all verified on simulator.

When a build white-screened, it diagnosed the Reanimated/New-Arch bug, fixed it, and re-verified itself.

This is the future of mobile dev. 🧵👇

[Lech Kalinowski @LeSiOO]

The magic isn't code generation — it's the loop.

plan → scaffold → pod install → build → run on iPhone sim → screenshot → judge its own output → refine → test → write SHIPPED.md

You set the visual bar. It does the 3am debugging.

[Michał Pierzchała @thymikee]

Say hi to 6 new models in our React Native Evals family:

#5 Claude Opus 4.8 → 84.7%
#11 Minimax M3 → 78.6%
#12 Kimi K2.6 → 77.7%
#14 Gemini 3.5 Flash → 75.5%
#17 MiMo V2.5 Pro → 73.4%
#23 Mistral Large 3 → 65.5%

Full benchmark ↓

[Lech Kalinowski @LeSiOO]

We've been building an on-device image-gen app driven entirely by the ambient-light sensor — no cloud, no prompts.

and we didn't just ship it, we measured it: every generated image traces back to the exact sensor reading, seed, and runtime path that produced it. fully reproducible, fully offline.

write-up coming 👀

[Daniel Williams @Danny_H_W]

Join us on June 11th in porto for @reactnativeOPO 2nd edition.

We’ll be covering developing with AI and adding ai to your apps.

You won’t want to miss it!

[Lech Kalinowski @LeSiOO]

@MonikaTarn68690 I know who was responsible for that! 😅. Probably no one foud out the hidden Easter Egg in background. Ther ia nice perk waiting to be found 😆.

[Lech Kalinowski @LeSiOO]

🪨 Caveman grunts. 🚀 Spaceman speaks.

[Lech Kalinowski @LeSiOO]

callstack/Apex thinks in React!

[Lech Kalinowski @LeSiOO]

“Most people overestimate what they can do in one year and underestimate what they can do in ten years.” — Bill Gates

[Callstack Engineers @callstackio]

Introducing Apex: a specialized, cost-efficient coding model for React Native.

Built by Callstack for the work our engineers do every day: building apps, fixing framework-specific issues, reviewing architecture, and shipping one codebase across platforms.

Now opening in private beta.

[Lech Kalinowski @LeSiOO]

Use it, save for later, leave a star... https://t.co/BxlZj096mE

[Lech Kalinowski @LeSiOO]

Two takes on the same problem: LLM tokens are expensive. Different ends of the pipe.

🪨 Caveman (by @juliusbrussee) makes Claude TALK like a caveman. Output-side compression — a Claude Code skill that cuts ~65% of output tokens. The model itself becomes terse. Failure mode: style drift if the model ignores the
skill.

🚀 SpaceSpeech (by @LeSiOO, @artus9033 ,@MonikaTarn68690 ) makes YOU talk like a spaceman. Input-side compression with local NLP only — spaCy + sentence-transformers + tiktoken. Zero extra LLM calls, fully deterministic, runs offline. Four strategies (Cadet →
Spaceman → Marine → Commander) auto-picked by token count × semantic similarity. Savings range 25–75% on input tokens. Failure mode: over-compression if you push the similarity threshold too low.

And it's no longer "compress and pray" — SpaceSpeech is now wired to Cloudflare AI Gateway, so you can fire the compressed prompt at Workers AI models (llama-3.1-8b, etc.) inside the same Win95-style UI and read the response
right there. Compress, test, ship.

Caveman shrinks the mouth. SpaceSpeech shrinks the prompt. Stack them and the LLM barely says a word in either direction.

🚀 https://t.co/vFzaw2PVXw

[Callstack Engineers @callstackio]

When the way software is built changes, we help companies adopt it.

For over a decade, that meant bringing React Native to production across mobile, web, desktop, and TV.

Now we are expanding that same mission to AI-native engineering: helping teams ship software with agents, safely and at scale.

This is our new manifesto.

[Michał Pierzchała @thymikee]

For a few weeks we've been building agents that are capable of understanding any app and then autonomously convert it to a React Native app.

Powered by agent-device and react-native-best-practices skills from @callstackio and some custom instructions

DM if you're interested

[Lech Kalinowski @LeSiOO]

I ve buid my own experimental Agency! 🤖
👇
https://t.co/gp0XWYAz0v

[Lech Kalinowski @LeSiOO]

Trying https://t.co/0xTqy31vH0 its freaking amazing... I don't like ZEP cloud hard dependency but still, free plan activated and cooking with local GPT OSS 120B. I am impressed!

[Lech Kalinowski @LeSiOO]

This update to My Chat Faces is personal.

My mom is losing her sight, and it made me realize how much AI tools still ignore readability.

So I built new accessibility features into My Chat Faces:

instant presets
Low Vision mode
Paper mode
better reply text sizing
stronger readability controls

It started as a way to make AI chats more beautiful.
Now it’s also about making them more humane.

Available in Chrome Web Store:

https://t.co/nQSEaryGJk

Open source:
https://t.co/1bRqRZ4VLD
https://t.co/9zgCs3L4tT

@callstackio

[AgentConf @AgentConf]

Call For Papers is now open. Submit your talk here 👇https://t.co/jq7moRnxnQ

We’re looking for bold, inspiring talks on AI engineering that will push the community forward.

Bring us your stories about:

🔹 vibe coding
🔹 agentic engineering
🔹 orchestration & multi-agent systems
🔹 adoption stories and real-world case studies
🔹 harnesses + benchmarks

What worked. What failed. What surprised you. What others should know before they build.

Selected speakers will join a lineup featuring voices from companies like Cognition, IBM, Expo and Rork.

We’ll cover your flight and accommodation.

You’ll share your ideas with 500+ engineers in Warsaw, building for the next shift in software.

Submission close: June 20

[Piotr Miłkowski @xeonbloomfield]

At @callstackio we're cooking something delicious!