AI engineer focused on agentic systems, applied ML research, and auditable production pipelines. Currently building healthcare-triage + multi-agent workflows, fine-tuning small LLMs (GRPO + LoRA), and shipping research code that re-runs end-to-end. I care about the reproducibility-vs-velocity trade-off — every artefact in this profile re-executes from a single command.
location: "🇸🇦 KAUST, Saudi Arabia ←→ 🇪🇬 Cairo, Egypt"
education: "MS Tech, Innovation & Entrepreneurship — KAUST"
working_on: ["agentic LLM workflows", "small-model fine-tuning (GRPO + LoRA)",
"multi-modal moderation", "applied research papers"]
publishing: ["NeurIPS 2026 — V-axis emotion centroids",
"NeurIPS 2026 — Cross-Architecture Substrate",
"INFOCOM / GSMA telecom-LLM track (×4 papers)"]
shipping: "EdGame — K-12 stealth-assessment learning games"|
End-to-end breast-cancer screening triage workflow that integrates a tuned RandomForest classifier (P3), a smolagents-style audit layer (P6), and a vendored CNN feature hook (P4). Zero malignant cases missed in the 114-case held-out cohort at the screening threshold.
|
Five-agent smolagents orchestrator-worker system for a paper-supply company: inventory +
quoting + sales + finance workers behind a customer-facing orchestrator. Wraps 7 SQLite
helpers as
|
|
ChromaDB-backed RAG agent over 15 video-game JSON records with LLM-as-judge retrieval evaluation and a Tavily web-search fallback. Three demo queries; the third correctly delegates to web search and cites a Wikipedia URL.
|
Reinforcement-learning fine-tune of Qwen2.5-3B-Instruct on a chain-of-thought letter-counting task using GRPO with LoRA adapters. Demonstrates training-time reward shaping on a small model.
|
|
CoT + ReAct travel-itinerary system. A Chain-of-Thought planner emits a strict
Pydantic-validated
|
Retrieval-augmented chat over Apollo 11, Apollo 13, and Challenger mission documents. Vector store + reranker + cited answers. Gradio UI; FastAPI backend.
|
|
Pipeline that moderates text, image, audio, and video with pydantic-ai, a streaming Gradio chat UI, and a FastAPI service layer. Structured outputs end-to-end.
|
Production K-12 ed-tech platform: 5 KAPLAY.js games, ECD (Evidence-Centered Design) analytics, 90K+ event samples. Built for a startup; live in classrooms.
|
Tip
Full project list — the Udacity AI Mastery Capstone spans 8 chapters (cd001-p1 … cd001-p8); the Agentic AI Nanodegree spans 4 (nd900-p1 … nd900-p4); research code lives under paper1_* … paper4_* and the vaxis-paper / substrate-paper NeurIPS submissions.
Languages
ML, deep learning, generative AI
Agentic AI & LLM tooling
Data & backend
Web, mobile, ed-tech
DevOps, infra, tooling
| Year | Venue | Title (repo / link) |
|---|---|---|
| 2026 | NeurIPS (under review) | Nine Emotion Centroids — A Label-Free Valence Axis Across Four Modalities |
| 2026 | NeurIPS (under review) | The Cross-Architecture Substrate |
| 2026 | Telecom-LLM track | Three Levers to Make LLMs Configure 5G Networks (catalog grounding + LoRA + RAG) |
| 2026 | INFOCOM track | Geometric V-Metric Instrumentation on Telecom Control Substrates |
| 2026 | Wireless control | Rate-Distortion Characterization of a 6-Bit VQ Codec (LLM + linear baselines) |
| 2026 | GSMA benchmarking | max_tokens × Prompt-Length Confound in Telecom-MCQ LLM Benchmarking |
| 2025 | Frontiers in Human Neuroscience | Stochasticity as a Solution for Overfitting (EEG inner-speech classification) |
| 2025 | Scientific Data | ArEEG — Arabic Inner Speech EEG Dataset |
| 2024 | EUROCAST | Symbolic Regression — Genetic Programming vs ML/DL |
| 2023 | 3ICT Conference | Smart Attendance Using BLE |
Reproducibility is a feature, not a constraint.