Abdullah AlJafari · CS at JUST · Open to AI/ML roles
My work spans computer vision, Arabic NLP/RAG, and MLOps—connecting models to tested APIs and deployment workflows.
Applied AI systems
Applied AI projects presented through the problem, implementation, and engineering decisions behind them.
Agentic RAG · Local-First AI
Answer questions about university Moodle courses from the student's own course files, with cited sources, on a local machine.
Result
Runs against a real Moodle account, backed by about 4,250 backend and 264 frontend tests.
Medical Computer Vision · MLOps
Classify chest X-rays as normal or abnormal through a deployable medical computer vision workflow.
Result
Hosted Streamlit demo returning a prediction with a Grad-CAM heatmap.
Arabic NLP · Sentiment Systems
Handle Jordanian and Levantine sentiment analysis where slang, sarcasm, and neutral sentiment routinely break generic Arabic classifiers.
Result
README-reported neutral recall of 78.5%, up from a 22.0% baseline.
More applied systems
Arabic NLP / RAG
Answer Arabic and English study questions from course materials instead of relying on generic model memory.
Result
Recorded grounded Arabic answers and honest abstention on out-of-scope questions.
Computer Vision · Detection & Tracking
Detect, classify, and track UAVs in RGB images and video through a full computer vision workflow.
Result
README video shows UAV detections with track IDs on an RGB drone clip.
Information Retrieval · RAG
Build a classical information retrieval system first, then extend it with grounded answer generation.
Result
README-reported MAP of 0.2976 and NDCG@20 of 0.3988 on Cranfield.
Computer Vision · Aerial Surveillance
Detect and track tiny vehicles and pedestrians in high-altitude UAV imagery where scale variance breaks standard detectors.
Result
Recorded dense aerial detection and Bot-SORT tracking on real traffic footage.
SKILLS SYSTEM
A concise view of the stack I use across model development, APIs, deployment, and product-facing interfaces.
PyTorch
training and serving vision models
OpenCV
preprocessing and postprocessing stages
YOLO / Object Detection
detection, tracking, and model comparison workflows
Object Tracking
ByteTrack and Bot-SORT multi-object tracking for UAV video
Hugging Face
Arabic model adaptation, demos, and model delivery
FAISS / Chroma / RAG
hybrid retrieval pipelines, Arabic normalization, and grounded responses
AraBERT
dialect-aware sentiment modeling for Arabic review data
LoRA / PEFT
parameter-efficient fine-tuning on local GPU hardware
LangChain
retrieval chains and hybrid search orchestration
Local LLMs
Ollama, Gemma, and LM Studio for local-first generation
Python
core language for training, data pipelines, and services
.png&w=48&q=75){kind=small}
FastAPI
typed model-serving APIs
Docker
reproducible runtime packaging
MLflow
experiment and artifact tracking
pytest / Ruff
automated tests and linting for ML codebases
Next.js / React
portfolio shell and product-facing interfaces
TypeScript / Tailwind
safer UI integration and fast styling systems
Streamlit
fast ML demos and experiment-facing interfaces
Azure / Deployment
cloud-ready hosting and delivery workflows
GitHub Actions
CI/CD automation and repeatable delivery
I study Computer Science at Jordan University of Science & Technology (JUST), focusing on applied AI systems. My work spans computer vision, Arabic NLP and retrieval, model serving, and MLOps — from training and evaluating models to building APIs and deployment pipelines.
I'm especially interested in systems where model quality is only part of the problem: small-object detection, Arabic-language edge cases, retrieval quality, inference constraints, and reliable behavior beyond a notebook.
Recent work includes a chest X-ray classification prototype with Grad-CAM, fine-tuning MARBERTv2 with LoRA for Arabic sentiment analysis, and building retrieval and vision systems around real deployment constraints. I use tools such as FastAPI, Docker, MLflow, and Azure to move projects beyond isolated model experiments.
Abdullah Aljafari
AI Engineer
Engineering journey
Dates are first commits in each public GitHub repository, or merge dates of this site's pull requests.
A DenseNet121 classifier behind a FastAPI backend and Streamlit frontend, with Grad-CAM explanations and automated deployment.
Jordanian Arabic sentiment analysis: MARBERTv2 adapted with LoRA and weighted loss, with a Hugging Face Spaces demo.
A bilingual Arabic/English RAG study assistant, and an aerial pipeline that uses SAHI slicing to detect tiny vehicles and pedestrians in UAV footage.
Drone detection and ByteTrack tracking with MLflow experiments, and a TF-IDF search engine evaluated on Cranfield before adding a RAG extension.
A local-first academic assistant over Moodle with hybrid retrieval, NLI-checked citations and approval-gated actions. The repository passed 780 commits by late October.
The evidence-led redesign went live on October 5, followed by Character Lab, a creative-coding playground, on October 6.
Exploration beyond the core stack
Smaller builds where I test one idea at a time: what I asked, what I tried, and what I learned.
Style 01/06·3D Toy
An experiment in AI-generated character consistency and browser-based image manipulation.
Six visual styles, real-time proportion controls, lighting and export — without regenerating the image on every interaction.
Style
Lighting
Computer Vision · Image Enhancement
How do low-light enhancement methods compare when improving visibility without ignoring artifact risk?
Finding
No single method wins everywhere, so each one is judged on visibility and on the artifacts it introduces.
Creative Engineering
Can a procedural 3D asset be presented without adding WebGL to the default page path?
Finding
The model loads only after a click, so the 3D preview adds no WebGL to the first page load.
Interactive Systems
How can a simple Flappy Bird-style mechanic become a fuller local multiplayer game system?
Finding
Two control paths, a health system and difficulty progression turn a one-button premise into a full local game loop.
Systems Research
What boundaries should a reusable AI orchestration pipeline define before implementation?
Status
Design notes only. Marked in progress on purpose instead of being shown as shipped work.
Verified credentials from DeepLearning.AI, Stanford, Cisco, Udacity, Udemy, Anthropic & IBM.
Let's build something useful
For hiring, project, or collaboration inquiries, use the form or email me directly.
I’m interested in internships and junior AI/ML roles involving vision pipelines, Arabic language retrieval, or model-serving APIs.