Rodion Kazennov | Nefayran

Software Architect | Tokyo 🇯🇵
Building in public
Japan Life Hub

[Rodion Kazennov | Nefayran @Nefayran]

AllKeep is on the web now:
https://t.co/NAEKwfoniQ

Take a photo of your stuff → the AI names and sorts it → ask where it is later instead of digging through boxes.

Started Android-only; now add and search from your browser too. The app also got multi-select, a home dashboard, and accessibility updates.

https://t.co/tjWjY0qIt7

[Rodion Kazennov | Nefayran @Nefayran]

@vladsiu https://t.co/NxRQTbtVmd

[Rodion Kazennov | Nefayran @Nefayran]

@ardent__dev https://t.co/NxRQTbtVmd

[Rodion Kazennov | Nefayran @Nefayran]

@foxtomb232 https://t.co/NxRQTbtVmd

[Rodion Kazennov | Nefayran @Nefayran]

Today I shipped my first app: AllKeep.

My wife and I had boxes everywhere and never knew what was in any of them. So I built something to fix it.

Snap a photo of your stuff — AI names it, sorts it, files it. Then just search "where's the drill" and it tells you.

https://t.co/tjWjY0qIt7

[Rodion Kazennov | Nefayran @Nefayran]

A small favor: I need 12 Android users to help me get AllKeep onto Google Play. Google requires 12 opted-in testers before I can publish.

https://t.co/4KRWJ3dbUX

If you're in, send me your Google account email (the one tied to your Play Store) and I'll add you to the tester list and send the opt-in link. Thank you.

[Rodion Kazennov | Nefayran @Nefayran]

Before sleep

[Rodion Kazennov | Nefayran @Nefayran]

AIエージェントが500回判断したのに、LLMを呼んだのは31回だけ。
パターンを学習したのかと思った。
実際はほとんど幻覚でした 🙂

小さいモデルの幻覚、みなさんどう対処してますか？

[Rodion Kazennov | Nefayran @Nefayran]

My AI agent made 500 decisions and called the LLM only 31 times.
I thought it had learned patterns.
Turns out it was mostly hallucinating 🙂

How do you deal with hallucinations in small models?

[Rodion Kazennov | Nefayran @Nefayran]

いろいろなLLMを使ってき���気づいたことがあります。

一番質の高いコードは、チャ���ト形式で使うと出やすい。
Opus 4.6 や Gemini 3.1 Pro は、大きなコンテキストを与えると特に強いです。

今、開発で一番困っていることは何ですか？

リプで教えてください。解決のヒントを返します。

[Rodion Kazennov | Nefayran @Nefayran]

After working with many LLMs, I’ve noticed something.

The highest-quality code usually comes from chat workflows.
Models like Opus 4.6 or Gemini 3.1 Pro perform best with large, rich context.

What’s the hardest problem you’re facing in development right now?

Reply, I’ll try to suggest a solution.

[Rodion Kazennov | Nefayran @Nefayran]

Google Cloud Tech innovations revealed → Are these really groundbreaking, or just more incremental updates? #googlecloud #tech

[Rodion Kazennov | Nefayran @Nefayran]

What do you think: Is the AI revolution just beginning, or are ethical challenges already too big to handle? 🤔 #ai #tech

[Rodion Kazennov | Nefayran @Nefayran]

Try it here: https://t.co/ubhSU05zGf
Know someone moving to Japan? Send them this.

[Rodion Kazennov | Nefayran @Nefayran]

Major update to Japan Life Hub 🇯🇵
AI Property Agent now covers 90 cities across Japan.
Tell it your budget and work location → get district recommendations with real rent estimates.
Here's what it does 🧵

[Rodion Kazennov | Nefayran @Nefayran]

Covers Tokyo, Osaka, Kyoto, Fukuoka — and 86 other cities.
6 languages: EN • 日本語 • 中文 • RU • हिंदी • ID
2 free credits to try it out.