Abdullah Al-Shobaki · AI Engineer

AI Engineer building real-world AI systems — from research to production.

Specializing in applied AI — from model to product, across vision, language, and real-time systems.

54% reduction in traffic wait times

AI-based adaptive traffic signal control · 0.949 mAP50

Introduction

I'm Abdullah — an AI Engineer who recently graduated with honors in AI Engineering from Bahçeşehir University (2022–2026). I've worked hands-on at the Apple Developer Academy | TUWAIQ, SCAI, and SDAIA, building computer vision, real-time, and multimodal AI systems end to end. I care less about a model hitting a benchmark and more about it holding up inside a real pipeline, with real latency and real failure cases. I'm looking for an AI/ML engineering role where I can take models from research into production.

Selected work

Four projects, chosen because each one shows a different part of the job: perception at real-time speed, interpretability, multimodal fusion, and shipping with a team.

01Flagship

AI-Based Traffic Management & Monitoring System

Adaptive signal control · Graduation flagship

54%

lower average wait time

0.949

mAP50 · 7 vehicle classes

12,064

synthetic training images

v1.0.0

tagged CI release

View source ↗

An end-to-end adaptive traffic signal control platform that replaces fixed-timer lights with AI-powered cameras. Perception, policy, and dashboard ship as one deployable system.

The problem

Fixed-timer signals cannot react to demand. Queues build on one approach while a green phase burns on an empty one — wasting fuel, time, and emergency response minutes. Cities need control that reads the road as it actually is, right now.

›View process & reasoning(6 steps)

01Framed the problem as perception → state estimation → policy, so each layer could be evaluated in isolation before integration.
02Built a 12,064-image synthetic dataset in CARLA across 13 weather presets to cover the long tail (night, rain, fog) that real footage rarely provides, then validated on real-world traffic camera video to check the sim-to-real gap.
03Fine-tuned YOLOv11m over larger variants after measuring the accuracy/latency tradeoff at real-time frame budgets.
04Added ByteTrack for identity-stable tracking, and OpenCV homography to map pixels to a road plane for true speed, queue length, and violation detection.
05Kept the policy layer pluggable — actuated, ALINEA, and a custom CompositePolicy — behind one interface, so control strategies could be swapped and benchmarked on identical traffic demand profiles.
06Wrapped it in a FastAPI WebSocket backend feeding a Next.js 14 dashboard, with a reproducible evaluation harness and Supabase Postgres storing every experiment run.

Stack

YOLOv11mByteTrackOpenCVPyTorchSUMOCARLAFastAPIWebSocketsNext.js 14Supabase PostgresDockerCI/CD

What I learned

The model was never the hard part — the evaluation harness was. Once every policy could be replayed against identical demand profiles, progress became measurable instead of anecdotal. Next time I'd build that harness on day one, and invest earlier in domain randomisation to narrow the sim-to-real gap before scaling the dataset.

Explainable AI for Autonomous Vehicles

Transparency for safety-critical decisions

sensor modalities fused

detector variants benchmarked

real-time

synchronized explanation panels

View source ↗

An end-to-end XAI system for autonomous driving in CARLA that makes every driving decision inspectable — visually and numerically — instead of leaving it inside a black box.

The problem

Autonomous driving stacks fail silently. When a vehicle brakes, swerves, or does nothing, engineers and reviewers need to know what the model actually attended to — otherwise safety review is guesswork.

›View process & reasoning(4 steps)

01Fused RGB camera and LiDAR into a single multi-modal perception pass so explanations reflect the full sensing picture, not just one stream.
02Benchmarked YOLOv5 Nano / Small / Medium against each other to find the accuracy point that still held real-time frame rates in simulation.
03Paired Grad-CAM (where the model looked) with SHAP (which features drove the decision) — visual and attribution evidence for the same moment.
04Synchronised LiDAR, attention maps, SHAP plots, and telemetry into one multi-panel real-time dashboard so explanations arrive with the decision, not after it.

Stack

YOLOv5PyTorchGrad-CAMSHAPLiDARCARLAOpenCVPython

What I learned

Explanations are a UX problem as much as a modelling one — a correct saliency map nobody can read at 20 FPS explains nothing. I'd next quantify explanation faithfulness rather than relying on visual plausibility.

Multimodal Chest X-ray Diagnosis

Vision Transformers + BERT

96.05%

fused F1-score

84.9%

ViT-only baseline

88.8%

BERT-only baseline

View source ↗

A diagnostic model that classifies 5 respiratory diseases by fusing chest X-ray imaging with the patient's written symptoms — the way a clinician cross-references image and history, rather than reading either in isolation.

The problem

Image-only models throw away the clinical context that radiologists rely on, and text-only triage misses what is visible on the film. Each modality alone plateaus well below what both together can reach.

›View process & reasoning(4 steps)

01Encoded the X-ray with a Vision Transformer and the symptom text with BERT, keeping both encoders independent so single-modality baselines stayed measurable.
02Fused the two representations through a learnable MLP classifier rather than naive concatenation, letting the model weight modalities per case.
03Ran ViT-only and BERT-only ablations to prove the fusion gain was real and not an artifact of extra capacity.
04Shipped a Gradio diagnostic app with attention heatmaps and token relevance, so a prediction always arrives with its supporting evidence.

Stack

Vision TransformerBERTPyTorchHugging FaceGradioPython

What I learned

Ablations earned more trust than the headline number. If I revisited it, I'd stress-test on missing or noisy symptom text — real intake notes are far messier than the dataset's.

Thalaja

Apple Developer Academy | TUWAIQ · First AI Foundation cohort

1st

AI Foundation cohort (competitive)

input modes: camera, voice, barcode

AR/EN

bilingual with full RTL

View source ↗

A shared, real-time family grocery app built in the first cohort of Apple's AI Foundation Program — camera, voice, and barcode all resolve to the same structured shopping list.

The problem

Household grocery lists live in scattered notes and chat messages. Typing items is the friction that kills the habit, and Arabic-speaking families are usually served last by tools like these.

›View process & reasoning(4 steps)

01Used CloudKit for real-time sync so every family member sees the same list without an account system to maintain.
02Added Gemini Vision for camera-based item recognition and barcode intelligence, so items enter the list by pointing rather than typing.
03Built voice-to-structured-items NLP so a spoken sentence becomes discrete, quantified list entries.
04Designed bilingual Arabic/English with full RTL support and Apple HIG conventions from the first screen, not as a retrofit.

Stack

SwiftSwiftUICloudKitGemini VisionCore MLApple HIG

What I learned

Shipping with a cross-functional team under mentorship taught me to defend technical decisions in plain language. RTL support proved cheap when planned early and expensive when deferred — a lesson I now apply to every interface I build.

Hackathons

Short-cycle builds — scoped, prototyped, and pitched under deadline.

Hackathon

Marsad (مَرْصَد)

AI League Hackathon 2025 · SCAI

AI Developer & Team Lead

AI-powered smart stadium platform for real-time crowd safety, fan engagement, and player performance analytics — YOLOv8 crowd detection with heatmaps for overcrowding and emergency alerts, MediaPipe pose estimation for player fatigue, and an Arabic NLP chatbot for fans. Aligned with Saudi Vision 2030 and positioned for FIFA World Cup 2034. Led development, technical implementation, and pitch delivery under hackathon deadlines.

PythonYOLOv8MediaPipeOpenCVTransformersNLP

Hackathon

Wadeh (واضح)

Allam Challenge 2024 Hackathon · SDAIA

AI Engineer

Selected from 200+ submitted ideas. An AI application that automates Arabic grammar analysis using advanced prompt engineering on IBM Watsonx with the ALLaM model — a multi-layered prompting strategy handles complex grammar across varying contexts without full fine-tuning. Designed with an accessible interface for educators, students, and language enthusiasts.

IBM WatsonxALLaMPrompt Engineering

Skills

Core

Where I go deepest

Computer VisionLLMs & Prompt EngineeringRAGPyTorchYOLOv8 / YOLOv11Multimodal AI (ViT, BERT)Real-Time Systems

Systems & Deployment

Getting models into production

FastAPINext.jsDockerMLflowAWS BedrockHugging FaceCI/CD

Foundations

Everything else I build on

PythonC++Swift / SwiftUISQLData Science tooling

How I work

Frame the problem

Read the domain before the dataset. Define what success measures, and what failure actually costs.

Prototype & validate

Train fast on real or synthetic data, benchmark variants against latency budgets, keep ablations honest.

Build the system

Wrap the model in what makes it usable — APIs, pipelines, dashboards, storage, and CI.

Measure & iterate

Reproducible harnesses over anecdotes. Every improvement has to show up as a number.

Recommendation

"I had the privilege of mentoring Abdullah during a 5-week AI Foundation Program at the Apple Developer Academy. Throughout the program, he demonstrated strong ambition, commitment, and an impressive ability to grasp new concepts quickly. His dedication to learning was evident, and he consistently approached challenges with curiosity and determination. In addition to his technical growth, Abdullah stood out for his collaborative spirit and willingness to work effectively with his peers. I am confident he will continue to excel in his journey and be an incredible asset to any team. I highly recommend him."

Shahad Alharbi · Software Engineer & Tech Mentor, Apple Developer Academy

Resume

The full record — education, roles, and publications.

One page, PDF, kept current.

Download resume ↓

Contact

Open to AI/ML engineering roles — always glad to talk about the work.

abdullah.hani2004@icloud.com

↗

linkedin.com/in/abdulahals