I build multimodal AI systems and agentic tools that stay inspectable when models, data, and environments become difficult.
01 · Practice
That means tracing the failure between components, building the interface that makes it visible, and turning the result into something a person can inspect, correct, and use.
02 · Featured work
Start with the outcome. Open the architecture, experiments, math, and failure analysis when you want the full story.
Generative AI · Visual systems
2026Controllable interior visualization across room staging, reference-guided editing, furniture retrieval, and 3D asset generation. I built the complete AI layer, the ComfyUI workflow system, Kaggle operations, and recommendation engine.
Document AI · Model & data engineering
2026A local document-understanding pipeline that detects tables, reconstructs their geometry, reads their text, and opens the result in a precise correction studio before export.
03 · Research systems
Research prototypes that expose their state: a concurrent embodied agent, a mathematical card engine, and a learning environment built around retrieval and authorship.
Concurrent agents
Fast planning, generated action, and continuous visual supervision on one interruptible event loop.
50 ms local reflex path · 3 s supervision cadence
Explainable game AI
Formal Basra rules, subset capture search, opponent inference, telemetry, and exact endgame reasoning.
Capture algebra · belief updates · 576 endgame pathsLearning protocol
A handwriting-first learning studio where the learner authors the note and the agent probes, checks, and connects it.
6 MCP tools · hybrid retrieval · reliable delivery states04 · Applied systems
Tools carried through interfaces, operating constraints, review paths, packaging, and real delivery.
Agentic security tooling
A constrained ReAct investigator gathers static PE evidence through nine tools and streams its trace live.
Inspect the agent ↗
Applied computer vision
A four-stage institutional photo quality gate with explicit accept, reject, and manual-review outcomes.
Follow the quality gate ↗
Arabic-first product engineering
One shared renderer adapted across desktop, Android, and Firefox with offline-aware prayer scheduling.
Explore three platforms ↗Eight implemented projects, from generative vision to local agents.
Browse every project05 · How I work
The same pattern appears across the work, whether the output is an image, table, action, report, or schedule.
Find the actual failure in the data, model, or environment.
Expose the intermediate state through a trace, editor, graph, or metric.
Adapt the model, constrain the agent, repair the data, or route to review.
Carry the result into a usable interface and preserve the full technical record.
06 · Background
Ain Shams University · 2021–2026
Agent P-DF document intelligence system
ArabicNLP 2026 participant
PDF · Experience, education, and technical record