Professional Profile
Ahmed Mustafa — AI Engineer | Machine Learning | Deep Learning | LLMs. Building intelligent systems with Machine Learning, LLMs and AI Agents.
I'm Ahmed Mustafa, a Software Engineer working in AI — I design and ship LLM-powered applications, AI agents and the data processing behind them, on production-grade engineering foundations.
Working Stack
What I Build
AI systems that do real work inside real products.
LLM Applications
Applications built around large language models — prompt design, scoring and evaluation, and token-aware cost control before anything reaches production.
AI Agents
Conversational agents that complete real tasks end to end rather than just answering questions — taking orders, handling enquiries, and driving flows to completion.
Data Processing
The pipelines underneath the model: cleaning, structuring and preparing data so predictions and dashboards work from reliable inputs.
AI Integrations
Adding AI capability to a product that already exists, through APIs and serverless functions, without rebuilding what already works.
About Me
I'm a Software Engineer who works in AI. That order matters: the models are one part of a system that also needs APIs, state, error handling and a deployment story — and that engineering half is where AI projects usually fail.
My AI work is shipped, not experimental. Prompt Score is a live tool for benchmarking and optimising LLM prompts, and the AI-Powered Restaurant SaaS runs a Gemini-backed agent that handles ordering and reservations over WhatsApp for real tenants.
How it fits together
- Software Engineering
- AI & Data
- Automation & AI Agents
- Domain Applications
10+
Mobile Apps
6+
Web Projects
4
SaaS Products
Core Toolkit
- AI & Machine Learning
- PythonMachine LearningDeep LearningLLMs
- Applied AI
- AI AgentsPrompt EngineeringLLM Integrations
- Engineering
- APIsData ProcessingNext.jsNode.js
Experience
The AI work sits on top of the same engineering practice — built, deployed and maintained as production products.
Full-Stack Engineer & SaaS Builder
DevIQ
Building end-to-end SaaS products, web platforms, and mobile apps with Next.js, NestJS & Flutter.
Software Engineer
Freelance / Remote
Developed web platforms with Next.js & NestJS, integrated APIs, databases, and CI/CD pipelines.
Mobile Developer
Freelance
Built 10+ cross-platform mobile apps with Flutter, Firebase, and clean architecture.
Featured Projects
Every project, ordered by relevance to the AI Engineer profile.
Prompt Score
Prompt Score is an innovative, web-based utility tool engineered to evaluate, benchmark, and optimize Artificial Intelligence prompts. It provides developers and prompt engineers with a structured, data-driven environment to refine their generative AI inputs, ensuring higher reliability, cost-efficiency, and predictable LLM outputs. 🛠️ Technical Tech Stack Frontend Architecture: Next.js / React (Deployed on Vercel) Styling & UI: Tailwind CSS (Optimized for modern, responsive dashboards) Localization: Built-in Internationalization (i18n) supporting multilingual interfaces (English/Arabic) API Integration: Serverless functions interacting with generative AI scoring models and evaluation metrics ⚡ Key Features & Engineering Impact Automated Prompt Benchmarking: Implements analytical scoring algorithms to grade prompt effectiveness, clarity, and structural constraints before deployment to production LLM pipelines. Token & Cost Optimization: Analyzes prompt length and token efficiency, helping developers reduce unnecessary overhead costs when executing large-scale API requests. Dynamic i18n Subsystem: Features a decoupled, configuration-driven localization architecture that handles seamless locale switching (/en and /ar) without breaking application state or layout. Micro-Frontend/Serverless Deployment: Fully optimized for low-latency delivery using Vercel's edge network, achieving near-instantaneous page loads and fast serverless execution. 🚦 Project Status Production Status: Live, fully operational, and accessible via web. Next Iteration: Integrating an automated multi-model sandbox to test and score a single prompt across various LLMs simultaneously.
Ather
Athar Category: Flutter App Tech stack: Flutter · Dart · Cubit · GetIt · gorouter · Kotlin · Shorebird Athar is a production Islamic companion app on Google Play, built around a single constraint that shaped almost every decision in it: it has to work with no internet, and it has to collect nothing about the person using it. Most apps in this category are thin clients. They fetch prayer times from a server, stream the Quran from someone else's CDN, and carry an analytics SDK that knows more about the user's day than the user does. Athar computes what it can, downloads only what it must, and ships no telemetry of any kind. The engineering below is mostly the cost of holding that line. The problem that set the direction The app originally fetched prayer times from a REST API. A user reported it plainly: "when I'm away from home, the prayer times don't work." The times worked on wifi and failed on a journey — which is precisely when someone most needs to know when to pray. Prayer times are solar geometry. They are computable to the minute from coordinates and a date, and needing a server for them was never anything but a dependency I had not removed yet. I moved the whole calculation on-device, with twelve calculation conventions (Umm al-Qura, Muslim World League, Karachi, Diyanet and others), both madhhabs, and per-prayer minute corrections so a user can align the app to the mosque they actually pray in. The calculation is pinned by tests against the published output of AlAdhan for six cities, two solstices and both madhhabs — because "I rewrote the maths and it looks about right" is not something to ship into someone's prayer times. Architecture Feature-first Clean Architecture across 25 features and roughly 430 Dart files, with data, domain and presentation separated per feature and dartz's Either carrying failures across boundaries instead of exceptions crossing them. Repository pattern over every source: prayer times, the Quran text and its page geometry, six tafsirs, the recitation catalogue, an endowment library, hadith, bookmarks, reading progress. Cubit for state, with locator-held singletons for anything that outlives a screen — the reader's position, the tasbeeh tally, the notification schedule — and per-screen instances where state must not leak between views. gorouter stateful shell so each tab keeps its own navigation stack. The native layer Eleven Kotlin classes, because parts of this cannot be done from Dart: An exact-alarm scheduler that arms several days ahead, re-arms itself when one fires, and rebuilds the chain after a reboot. Android's Doze mode batches inexact alarms, which is why reminders used to arrive four at a time an hour late — a bug that only exists on a real device, left idle, overnight. A foreground media service for the adhan, so it plays with lock-screen controls and can be stopped from the notification shade. Three home screen widgets — prayer times, an ayah, a du'a. A vibration channel. Flutter's HapticFeedback routes through View.performHapticFeedback, which Android silently drops whenever the system's touch-feedback setting is off — the default on several skins. The tasbeeh counter buzzed on some phones and not others, and raising the impact strength changed nothing, because the strength was never the problem: the call was not arriving. Talking to the vibrator directly fixed it. Size, and what it costs to be complete The Madinah mushaf renders from 604 fonts — one per page, about 67 MB. Bundling them is the obvious approach and it puts 67 MB into every install, including for the majority who never open the mushaf view. They are fetched per page on first view and cached, alongside the same treatment for tafsirs, hadith and the library. The user's download went from 85 MB to 21 MB, and the app still contains all of it. The largest of those is an endowment library of 63,433 books, browsable and readable in-app, downloaded per book with a checksum — because a truncated download is still a file, and without a digest it sits on the shelf looking installed and opens to half a book. Privacy as an enforced constraint No Firebase. No analytics. No crash reporting. No advertising SDK. No account, no login, no server of my own that user data could reach. This is enforced rather than intended: cleartext traffic is disabled app-wide with a single documented exception, every URL parser rejects a non-https scheme, and a test reads the published privacy policy and checks it against the app's own constants — so the policy cannot drift away from what the code does. It had drifted once: the policy still described sending coordinates to a third party months after the calculation moved on-device, which is an overstatement, but a policy that does not describe the app is a broken policy in either direction. What it does Quran — the Madinah mushaf, ayah-by-ayah reading, six attributed tafsirs, 241 reciters across 20 riwayat, verse-level audio timing that highlights the ayah as it is recited, bookmarks, khatma plans, and a memorisation mode that hides the verse until you have recited it. Prayer & adhan — on-device times, a month table, twelve conventions, per-prayer sounds, a pre-prayer notice, an iqama offset, and a sahih hadith attached to every prayer notification with its grade and source. Library — 63,433 books, searchable, with multi-volume works handled as one download. Daily — azkar with a configurable typeface and size, du'a, the full ruqya, hadith, seerah, prophets' stories, a Ramadan section, qibla by compass, a tasbeeh counter, a deeds tracker, user-written reminders, and Quran radio. Engineering practice 731 tests, and the interesting thing about them is not the number. Most were written against one specific failure, and several assert that the bug still reproduces without the fix — so the test cannot quietly pass for its own reasons. Some are structural, and each exists because that class of bug shipped once: one reads every assets/ path in the source and demands pubspec.yaml declare it — after a JSON file sat in the repo undeclared for months, so a whole feature had never once worked; one rejects any context.read() in a screen that cannot reach a provider — after a settings toggle threw on every tap, saved the preference, swallowed the error, and did nothing; one refuses to offer a font in a picker that the app does not actually bundle — after two fonts were referenced, credited and licensed without ever being shipped. Every release carries a changelog entry, enforced by a test, and written for a user of the last published build rather than for me — a fix to something no one has run yet is development history, not news. Delivery uses Shorebird code push for Dart-only fixes and Play In-App Updates for the rest, so a correctness bug can reach users in minutes instead of a review cycle. Status Published on Google Play and in active development. The store build is an early version; 42 versions of work sit ahead of it — offline prayer times, the library, the media catalogue, the native alarm layer, the support page — now going through a device-testing and bug-fixing cycle before a single large release. Next: publishing that release, then a licensed adhan recording set, and a recitation-checking mode using on-device speech recognition.
Bella Cosmetics
Bella Cosmetics Category: E-commerce Web Application Focus: Online Store · Product Management · Shopping Cart · WhatsApp Ordering · Responsive UI Bella Cosmetics is a modern e-commerce platform built for a natural personal-care brand, providing customers with a simple way to discover products, browse categories, view offers, and prepare their purchases for ordering. The platform is designed around a straightforward commercial goal: turn a product catalogue into a simple purchasing experience without unnecessary complexity. The problem Small beauty and personal-care businesses often depend heavily on social media and direct messaging to display products and receive orders. That creates several problems: Products are difficult to browse systematically. Prices and offers can become scattered across posts. Customers have no central product catalogue. Finding a specific product takes unnecessary effort. The business has limited control over how its products are presented. Bella Cosmetics was built to provide a dedicated digital storefront where products, categories, pricing, offers, and ordering are organized in one place. Product catalogue The store organizes products into clear categories, including: Skin Care Hair Care The current catalogue includes products such as hyaluronic-acid serum, vitamin C serum, shampoo, body butter, natural oil mixes, lip balm, deodorant cream, and body splash products. Each product is presented with its category, name, description/details where available, and price. Offers & promotions The homepage includes a dedicated offers section for products with limited-time discounts. For example, the current store highlights a hair-care cream with a discounted price compared with its original price. This creates a dedicated commercial space for promotions instead of relying on social-media posts alone. Shopping experience The application provides a shopping-cart experience that allows users to add products while browsing and review their selected items before proceeding with the order. The interface is intentionally simple: the customer can move from: Browse → Product → Cart → Order without going through unnecessary steps. WhatsApp ordering Instead of forcing the business into a complicated payment and shipping infrastructure, the platform integrates the existing customer communication workflow through WhatsApp. Customers can use the WhatsApp contact option to communicate with the business regarding orders and inquiries. This approach is particularly suitable for a growing business that wants an online storefront while keeping order confirmation and customer communication human and flexible. Content & brand experience The website also communicates the brand story rather than functioning as a product grid only. The "Our Story" section presents Bella Cosmetics as a handmade personal-care brand focused on natural and organic ingredients and describes the quality-control process used for its products. This gives the storefront a stronger connection between product discovery and brand identity. Responsive experience The interface is designed as a responsive web experience so customers can browse the catalogue, view products, and manage their cart across different screen sizes. The structure prioritizes: Clear product discovery Simple navigation Category filtering Visible pricing Promotional products Shopping cart access Direct customer communication What I built The project involved developing the storefront experience and connecting the different parts of the shopping journey: Home → Categories → Products → Product Details → Cart → WhatsApp The goal was not simply to create a visually attractive cosmetics website, but to build a practical digital storefront that could support an actual small-business sales workflow. Business outcome Bella Cosmetics transforms a traditional product catalogue into a structured digital storefront where customers can: Discover products. Browse by category. Compare prices. Find current offers. Add products to a cart. Contact the business directly through WhatsApp. The architecture also leaves room for future expansion into more advanced e-commerce capabilities such as online payments, order management, customer accounts, inventory management, delivery tracking, and analytics.
Torjoman
Torjoman is a cross-platform enterprise ecosystem engineered to fully digitize organization workflows, management, and live broadcasting. This end-to-end solution bridges the gap between administrators, field managers, speakers, and end-users by orchestrating a production-ready backend with 5 dedicated web and mobile applications. Ecosystem Tech Stack Backend Core: NestJS v11 (Monorepo Architecture) | PostgreSQL 16 | Prisma ORM Context & Security: nestjs-cls (AsyncLocalStorage) | JWT | Global Guards Real-time & Delivery: Socket.io Gateways | Production-Ready Docker Infrastructure Client Apps (Web & Mobile): Flutter | Clean Architecture | BLoC / Cubit State Management | Shared UI Core (torjomancore) The 5-App Ecosystem Architecture The system is engineered as a unified codebase that powers 5 distinct client-side applications tailored for specific user roles: Torjoman Owner (Web): Global platform dashboard for SaaS subscription tracking, resource monitoring, and multi-organization provisioning. Torjoman Admin (Web): Super Admin control panel for single organizations to manage internal roles, managers, and regional settings. Torjoman Manager (Mobile/Web): Operational app for field managers to handle real-time location mapping (coordinates), site-specific setups, and speaker assignments. Torjoman Studio (Mobile): A dedicated mobile streaming app for speakers to initiate live sessions, broadcast audio, and generate on-the-fly transcriptions. Torjoman Live (Mobile/Web): The consumer-facing application allowing users to discover live locations, stream real-time audio, read live transcriptions, and save favorites. Key Engineering & Architectural Solutions Hybrid Multi-Tenancy Engine: Built to natively support both Shared-DB SaaS Mode (commercial multi-tenancy) and Enterprise Dedicated Deployments (private on-premise installation for government ministries) without codebase forking. Automated Data Isolation Layer: Integrated a custom PrismaQueryExtension that automatically intercepts all database transactions to inject organizationId filters dynamically. This guarantees hard logical data isolation between clients and eliminates human-error leaks. Edge Speech-to-Text (STT) Pipeline: Utilizes Device-Side Speech-to-Text on the mobile client apps to process live speech directly into text. This architectural choice bypasses massive audio server storage costs and drastically optimizes network bandwidth. Dynamic Regional Mapping: Supports optional hierarchical data partitioning (useRegions) that can be activated for large-scale state structures or disabled for flat corporate clients. Current Status Core & Network: 100% operational across the multi-tenant backend architecture, Prisma isolation layers, and cross-app communication protocols. Deployment: Fully dockerized, security-hardened, and optimized for end-to-end production environments.
Remontada
Remontada is an interactive, gamified sports trivia application built for a client via Upwork using Flutter. Inspired by popular football challenge shows on YouTube, the game tests users' sports knowledge by displaying dynamic player visual cards alongside contextual textual hints, requiring opponents or single players to strategically guess and reveal the correct football player. 🛠️ Technical Tech Stack Framework: Flutter (Dart) State Management: BLoC / Cubit Pattern (To strictly decouple game progression, timers, and score scoring logic) Architecture: Modular, Clean Architecture (Tailored for seamless addition of new quiz packs) Assets & Performance: High-performance local caching for image assets to minimize rendering latency during fast-paced rounds ⚡ Key Features & Engineering Architecture Dynamic Game State Controller: Engineered a highly responsive state machine using BLoC to handle real-time game flows, managing active rounds, score multiplication, instant feedback loops for correct/incorrect answers, and transition timers. Hint Injection & Image Masking Subsystem: Developed an asynchronous widget structure that seamlessly pairs high-resolution player graphics with specific textual hint strings, managing asset loading smoothly to maintain a constant 60 FPS. Scalable Content Architecture: Structured the codebase utilizing abstract data repositories, allowing the client to easily scale the game by injecting new player packages, seasonal challenges, or multi-category questions via local configuration files. Production-Grade Delivery: Successfully delivered the project on an agile timeline under Upwork's client specifications, ensuring zero UI memory leaks and achieving full cross-platform responsiveness across both iOS and Android platforms. 🚦 Project Status Core Status: 100% Completed, verified, and delivered to the Upwork client. Core quiz pipelines, scoring systems, and asset rendering engines are production-ready.
YouTube Channel
Join the community! I regularly share tutorials, tips, and deep dives into software engineering, Flutter, Next.js, and more.
Discuss an AI Project
Working on something AI-shaped? Tell me the problem and I'll tell you whether a model is the right answer.