How I build AI systems
Most of my work sits in financial operations, real estate, and document processing — domains where a wrong answer has real cost, so the system has to be designed around that.
A pattern runs through that work: the model may propose, and the system decides what may commit. Deterministic code handles anything that is checkable. Evaluation is infrastructure, not a phase after the demo.
01
Constrained writes
Models propose structured intents. Humans or rules confirm before anything commits.
02
Document intelligence
Vision and staged pipelines for statements and documents that have no fixed layout.
03
Evaluation harnesses
Rule-based checks, model-graded metrics, and persona regression to catch drift.
04
LLM orchestration
Multi-service agents with phase gates, retrieval, and thin channel adapters.
Selected systems
Production AI systems I designed and built — problem, constraints, architecture, and the decisions I would defend.
Research & Publications
Peer-reviewed work in brain-computer interfaces — stated carefully against the paper itself.
NEUROPHONE: Real-Time Brain-Mobile Phone Interface
A P300-based brain-computer interface that lets people with motor disabilities operate a smartphone through visual attention alone. I contributed to EEG data collection, signal preprocessing and CNN training.
98%
Average accuracy, 5-fold CV
0.95
F1 score, 5-fold CV
A lightweight CNN reached 98% average classification accuracy and a 0.95 F1 score under 5-fold cross-validation, ahead of EEGNet, ChronoNet, DCRNN and RNN baselines on both the EPFL BCI dataset and our own recordings.
The in-house dataset was small and the end-to-end delay between phone and inference API was never accurately measured. Selecting an icon took three to four flash repetitions, roughly eight seconds.
Where I have built
Roles and the systems behind them — co-founding Youtiva, independent AI work, and earlier research internship experience.
Co-Founder & AI Engineer
Youtiva
Own AI architecture across the products we build, and write the systems myself. The work spans conversational systems over business data, document extraction pipelines, retrieval, evaluation, and the backend services that carry them.
- Built a conversational accounting platform for small businesses covering chart-of-accounts generation, transaction classification, invoicing and bank reconciliation.
- Designed a confirm-then-commit pending-action queue so the model proposes ledger mutations but cannot execute them without explicit human confirmation.
- Built an AI buyer's agent for residential real estate running a staged journey from discovery through to generated offer documents, dispatching to backing services for property records, comparables and permits.
- Established the evaluation harness — rule-based metrics alongside model-graded ones, plus replayable persona regression tests to catch behavioural drift from prompt and model changes.
- Delivered one AI capability across web chat, streaming video avatar, voice and telephony by keeping orchestration central and channel services thin.
- Authored the team's backend and frontend engineering standards, including a shared streaming contract and API response envelope.
AI/ML Engineer
Independent
Top Rated on Upwork with a 100% job success rate. Natural-language interfaces over business data, predictive modelling, and computer vision for international clients.
- Delivered AI and machine learning projects spanning natural-language interfaces over business data, predictive modelling, and computer vision.
- Built a natural-language analytics service that generates and executes pandas code against client datasets to answer business questions.
AI Engineer, Intern
RadicalX AI
Developed and optimized models for career-coaching and integrity-monitoring products using OpenAI APIs and TensorFlow.
- Worked on career-coaching and integrity-monitoring products using OpenAI APIs and TensorFlow.
Bachelor of Computer Science
Ain Shams University
Scientific Computing, Faculty of Computer and Information Science · Cairo, Egypt · 2019 — 2023
Technical stack
AI systems
Evaluation
Backend
Data and retrieval
Infrastructure
What Clients Say
Independent client feedback across AI, machine learning, automation, and software projects.
5.0
Average Rating
100%
Job Success
Top Rated
Upwork
Build an ML price prediction model
“Great work and great person very happy to work with her always.”
Python Developer for LinkedIn Content Automation
“Highly competent and demonstrates clear expertise in her work!”
AI / Machine learning and IoT task
“Exceptional and clearly knows what she's doing!”
Transformation of points and mapping coordinates - Python
“There's no question about it — truly a master at delivering exceptional work!”
Founder
“An outstanding AI engineer with exceptional technical skills and a deep understanding of AI and machine learning. Consistently delivers high-quality work, solving complex problems with efficiency and creativity. A great team player with excellent communication skills, always willing to share knowledge and mentor others.”
Machine Learning Expert for Educational Project
“Working with her was an absolute delight! She delivered exceptional results with professionalism and expertise. Highly recommended!”
More context on Upwork.
Notes from building
Patterns that keep showing up when AI systems move from demo to production.
Let's talk about applied AI
I enjoy comparing notes on production AI systems, evaluation, document intelligence, retrieval, and reliable automation.