Sayed Ali Alkamel

👨‍💻 Flutter GDE | 🌍 Founder @FlutterMENA | 🚀 Building apps & breaking limits | 🎤 Storyteller in tech | 🇧🇭 Bahraini | Tech with a human touch ✨

[Sayed Ali Alkamel @Sayed3li97]

I'm glad to announce that I'm officially a @GoogleDevExpert for @Dart and @FlutterDev

I am thankful for the amazing Flutter community and I look forward to contributing more to this community🎯

#flutter

[Sayed Ali Alkamel @Sayed3li97]

Google shipped a knowledge spec for AI agents in June. The whole thing:

- a folder of markdown files
- one file per table, metric, API, runbook
- YAML frontmatter on top
- exactly one required field

No SDK. No account. It is called OKF.

https://t.co/jog4Kfh4tZ

[Sayed Ali Alkamel @Sayed3li97]

Two things parallel agents are good for that people forget:

1. Speed. Independent subtasks shouldn't run one at a time.
2. Confidence. Ask the same question 3x and vote.

Same pattern, two very different jobs.

https://t.co/xwcfSmPiOL

[Sayed Ali Alkamel @Sayed3li97]

A single prompt that handles every request is the AI version of a 2000-line function.

Routing is the refactor: classify the input, send it to a small prompt built for that case.

Separation of concerns, but for prompts.

https://t.co/kYeZMmLdTH

[Sayed Ali Alkamel @Sayed3li97]

The cheapest reliability win in LLM apps:

Stop making one prompt do five things.

Split it into a chain. One job per step. A bit more latency, a lot more accuracy, and you can see exactly where it breaks.

https://t.co/VWA6RdUsLg 

[Sayed Ali Alkamel @Sayed3li97]

The augmented LLM is the base unit of every agent: one model plus retrieval, tools, and memory. When to use it, and when a single agent is enough. https://t.co/98kH0TXle9

[Sayed Ali Alkamel @Sayed3li97]

Most "AI agents" are just an LLM with retrieval, tools, and memory.

Not an insult. Usually the right call.

Ship the simple version first. Add autonomy only when you have to.

https://t.co/98kH0TWNoB

[Sayed Ali Alkamel @Sayed3li97]

Flutter calling C directly. No platform channel, no serialization, just a synchronous function call.

That is Dart FFI, and I wrote up what it is, when to actually reach for it, and the gotchas I hit building camera_pro (shared C core, pixel work on every frame).

for more https://t.co/ctCWRPr0uT

[Sayed Ali Alkamel @Sayed3li97]

"The LLM Preamble Problem: How RLHF Made Your Model Too Polite to Ship" by @Sayed3li97 #DEVCommunity https://t.co/Tl3iO0x0Jv

[Sayed Ali Alkamel @Sayed3li97]

Your API gateway treats the request body as a sealed envelope: reads the address, forwards the letter. An AI gateway opens it. It meters tokens, caches by meaning, swaps models on failure. Same shape, almost nothing else in common.

https://t.co/cvUcPGxrHs

[Sayed Ali Alkamel @Sayed3li97]

Is this the future of Flutter?

This app has zero pre-designed lesson screens. Gemini reads what the learner struggles with and assembles bespoke widgets in under 2 seconds. Home, vocab carousel, root explorer, quiz. All composed at runtime.

Full tutorial 👇

https://t.co/kAvZJl3DmD

#Flutter #AI #GoogleIO2026

[Sayed Ali Alkamel @Sayed3li97]

I wrote up how Kalaam works - the full technical picture.

How a Gemini model composes a Flutter widget tree at runtime. How the Root System Explorer, Live Inspector, and 13 custom Arabic widgets plug into the GenUI SDK. Why bidirectional loop + data model binding changes what "lesson" even means.

https://t.co/yGdIAUDAA2

And yes, it's all open source:
https://t.co/DTquMi8KCx

[Sayed Ali Alkamel @Sayed3li97]

Type a goal. Gemini builds the lesson.

"Ordering coffee" → scene-setter, root diagram, vocab cards, pronunciation drill, role-play — sequenced live, not scripted.

Or pick a real-world scenario. Either way, the interface is assembled at runtime by a model that knows Arabic pedagogy.

Clone → flutter run --dart-define=KALAAM_DEMO=true → lesson in 30 seconds. No Firebase, no API key.

https://t.co/bO12wrcW42
#Flutter #GenUI #OpenSource #Arabic

[Sayed Ali Alkamel @Sayed3li97]

Tap {} and watch Gemini write your UI in real time.

Every widget you see on screen started as a JSON message. The Live GenUI Inspector streams it raw — field by field, as tokens arrive — so you can literally watch "this JSON became that widget."

CREATE turn_2 → Root Explorer with ش-ر-ب.
"transliteration": "sukkar" → VocabCard.
No magic. Just a model, a catalog, and a wire format.

Kalaam · كلام — open source Arabic tutor built on Flutter GenUI
https://t.co/DTquMi8KCx

[Sayed Ali Alkamel @Sayed3li97]

Every Arabic word grows from a 3-letter root.

Tap شَرِبَ (he drank) on the Root System and it blooms open: وزن مَفْعُول, transliteration, an Explore→ button that asks Gemini to go deeper.

The diagram isn't static. It's a branch point. Kalaam builds this UI at runtime from a Gemini model prompt.

Open source. Run it in 30 seconds with no API key.

https://t.co/bO12wrdtTA

[Sayed Ali Alkamel @Sayed3li97]

Full article: https://t.co/RryF9qmHHp

[Sayed Ali Alkamel @Sayed3li97]

Evaluated Spring AI for a fintech. The number I can’t shake: 74% of Java devs use AI tools. Only 34% build AI-powered apps with any framework. Is the Python sidecar still the default in enterprise Spring Boot shops?

[Sayed Ali Alkamel @Sayed3li97]

Hot take: token usage will be this decade's "lines of code" metric.

Everyone will add it to dashboards. Teams will optimize for it. It will tell us nothing.

What metric do YOU think actually captures dev productivity in the AI era?

https://t.co/xKV8ErkTAV

[Sayed Ali Alkamel @Sayed3li97]

Two operating modes in agentic engineering. Most developers use both in a single day without ever naming them.

𝗖𝗼𝗻𝗱𝘂𝗰𝘁𝗼𝗿 𝗺𝗼𝗱𝗲 Real-time, hands-on direction. You are in the IDE, watching code appear, guiding with prompts and corrections at each step. Best for: exploratory work, novel problems, tasks where the output cannot be fully specified in advance. Ceiling: if you personally direct every interaction, throughput improvement stays capped.

𝗢𝗿𝗰𝗵𝗲𝘀𝘁𝗿𝗮𝘁𝗼𝗿 𝗺𝗼𝗱𝗲 You define goals and assign them to agents. You review results without watching code appear line by line. Best for: well-defined tasks, bug fixes against established patterns, test generation, codebase migrations. Ceiling: requires a precise spec before you can delegate. You must write it before the agent can execute it.

The honest part that most content on this topic skips: 𝗼𝗿𝗰𝗵𝗲𝘀𝘁𝗿𝗮𝘁𝗼𝗿 𝗺𝗼𝗱𝗲 demands more upfront discipline, not less. Teams under sprint pressure will find this counterintuitive. The payoff comes in the second sprint, not the first.
Knowing which mode the task warrants, and switching deliberately, is the skill.

Full framework with more detail on when each mode breaks down, including specific tooling examples: https://t.co/ar46WtkBJP

hashtag#AgenticEngineering hashtag#SoftwareEngineering hashtag#Flutter

[Sayed Ali Alkamel @Sayed3li97]

3 numbers every engineering leader should bookmark:

📊 861% code churn under heavy AI (Faros AI)
📊 −19% speed with top AI tools (METR, 2025)
📊 9.4x higher churn for regular AI users (GitClear)

More tokens ≠ more value shipped.

https://t.co/xKV8Erlrqt

[Sayed Ali Alkamel @Sayed3li97]

Full article, interactive BPE animation, provider comparison table, and working token counting code:

https://t.co/U8qPoRC7ve

[Sayed Ali Alkamel @Sayed3li97]

5 rules for working with tokens in production:

1. 1 token = ~4 English chars or 0.75 words
2. Arabic, Chinese, Thai: 2-3x more tokens per word
3. Code and prose tokenize completely differently
4. System prompts and tool schemas count too (most devs undercount by 30-40%)
5. Context quality beats context size

Save this.

[Sayed Ali Alkamel @Sayed3li97]

Tokens are not a billing detail.

They decide whether your RAG pipeline fits in one request, whether your agent loop runs 5 turns or 50, and whether your bill is $200 or $2,000 at scale.

A compression trick from 1994 is still doing one thing: making redundant patterns smaller.

Full breakdown with BPE animation and code 👇