Hi, I’m AlaaAl‑bdewi

02—How I work

The rules I work by.

Short, testable, and learned in production. Each one comes with a number or a note I wrote about it.

01

Judge it by the 100th run, not the first

One great answer doesn’t make a prompt finished. It makes it lucky. I call a prompt or agent tuned when about 98% of results meet my spec over many runs with real inputs, on the same model. Then I try a smaller, cheaper model. If it still hits 85–92%, it’s robust.

From Production Notes #01

02

Numbers beat adjectives

“Much cheaper” is an opinion. “82% lower generation cost” is a result, and it came from three changes together: routing each step to the smallest model that passes its eval, caching stable prompt prefixes and running steps in parallel. Quality and speed get numbers too: 91% on our internal rubric, first output in 15–30 seconds.

From Case study · Muqnea AI pipeline

03

Work in the language you can verify

I prompt in the language I can check the answer in. The risk isn’t the prompt, it’s the reply: a plan full of half-understood terms that you accept anyway. I hold agents to the same rule. They hand off schema-validated structured output, not free text, so every step can be checked.

From Production Notes #02

04

Fit the system, not my habits

In an existing codebase, I follow the conventions already in place instead of adding my own. On Minaret, I worked on the shared core of an ERP platform, so every change had to hold across the whole platform, not just the screen in front of me.

From Case study · Minaret ERP

03—What I do

I own systems end to end.

Back end, data and front end, plus the multi-agent AI layer that now sits on top of them.

Aai

AI systems in production

Multi-agent LLM pipelines that run every day, with cost, speed and quality I can measure.

• →Specialised agents with schema-validated handoffs
• →Model routing, prompt caching, parallel steps
• →Eval rubrics and output scoring
• →RAG, streaming and LLM API integration

Bbackend

Backend & APIs

Services and APIs that the product and other services build on.

• →Python microservices and queue-backed workers
• →Versioned REST APIs with auth and rate limiting
• →Laravel (incl. HMVC), PHP 8 and Node.js
• →Stateless services, built to scale

Cdata

Data at scale

Data models that match how the business works, and queries that stay fast at millions of rows.

• →Relational modelling from real business processes
• →Composite indexes, partitioning, query rewrites
• →Stored procedures for heavy reporting
• →PostgreSQL, MySQL and Redis caching

Dfrontend

Front ends people use

Interfaces for real users and roles, kept fast as the data grows.

• →Component-driven SPAs in Vue.js and React
• →Role-scoped dashboards and admin screens
• →Pagination, virtualised lists, client-side caching
• →Response contracts shaped with the backend team

04—Over coffee

Pull up a chair.

The questions people usually ask me, answered the way I’d answer them over coffee.

05—The stack

From UI to AI.

One system, five layers. This is what I use at each one.

UI01 / 05

Frontend

Interfaces people use every day, kept fast as the data grows.

• Vue.js
• React
• Next.js
• TypeScript
• JavaScript (ES6+)
• Tailwind CSS
• SPA and component-driven architecture
• State management
• HTML5 / CSS3

API02 / 05

Backend

Services, APIs and workers that products and other services rely on.

• Python
• PHP 8
• Laravel (incl. HMVC)
• Node.js
• REST API design
• Microservices
• Message queues and background workers
• Stored procedures
• SOLID

DATA03 / 05

Data

Data models that match how the business works, and queries that stay fast at multi-million-row scale.

• PostgreSQL
• MySQL
• Redis
• Relational modelling
• Composite indexing
• Partitioning
• Query optimisation
• Caching strategies
• Pre-aggregated reporting

AI04 / 05

AI & LLM engineering

Where most of my work is now: agents that hand off structured output, with cost and quality I can measure.

• Multi-agent systems and orchestration
• Prompt engineering with schema-validated structured output
• LLM API integration
• Retrieval-augmented generation (RAG)
• Model routing and token-cost optimisation
• Prompt caching
• Streaming responses
• Evaluation rubrics and output scoring

OPS05 / 05

Platform & DevOps

The tools that take code from my machine to production.

• Docker
• Git / GitHub
• CI/CD
• Linux
• Nginx
• AWS

The full stack

06—Journey

From web apps to AI systems.

Five years in production: small-business apps, an ERP core, solo systems for real clients, and now multi-agent AI.

01January 2021 – January 2022

Junior Freelance Developer

Independent

I built web apps for small businesses end to end, in PHP, Laravel, Vue.js and MySQL.

02August 2022 – March 2024

Full-Stack Developer

Minaret · AL-MOTKAMEL Global Co.

I worked on the shared core of Minaret, a complete ERP platform, at multi-million-row scale — and led the Akar redesign.

03March 2024 – May 2024

Full-Stack Developer

Be In Media

I took over an HR and payroll module that gave wrong results and brought it to a stable state.

04May 2024 – May 2025

Freelance Web Developer

Independent

I delivered three production systems solo — clinic, transport and an enterprise CMS — from requirements to support.

3production systems, delivered solo

05August 2025 – August 2026

Frontend Developer

Pazar Pro · IRIS

I led the front end of a free-zone multi-vendor marketplace (B2B, B2C, B2B2C) through to handover.

NowMay 2025 – Present

Backend & AI Systems Engineer

Muqnea

I own the backend and multi-agent AI layer of a live generation platform — 82% lower cost from model routing, caching and parallel steps.

82%lower generation cost

Full experience

07—Selected work

Systems people actually run.

Case studies told with the numbers I can share.

08—Production Notes

I write down what works.

Building AI in production. English and Arabic, twice a week.

PRODUCTION NOTES · #01PROMPT & AGENTS

Lucky ≠ Tuned

A prompt that gives you one great answer isn’t finished. It’s just lucky. Here’s how I tell the difference.

Read the noteOn LinkedIn

PRODUCTION NOTES · #02ARABIC × AI

Prompt in the language you can check

The best language to prompt AI in isn’t always English. It’s the language you can catch its mistakes in.

Read the noteOn LinkedIn