Twenty-seven works, hung across five rooms. Every piece here has a hard core built from scratch — an algorithm, a data structure, a training loop. A pretrained model is never the whole work; it is one material among several, and the difficult parts are made by hand.
Visitor information
On understanding what is underneath before building on top of it.
Hesham Ahmed Wageeh Elsayed is an AI engineer based in Cairo. The throughline of his work is simple: understand what is underneath before you build on top of it. He writes the hard parts himself — the parts most people import — because knowing how an algorithm actually works is the difference between using a tool and engineering one.
That instinct became a company. MalwareGuard, hung in the Founder’s Room, is an AI-powered cybersecurity platform he designed and built end to end: real-time threat detection combining deep learning with NLP behavioural analysis, and an LLM layer that writes the threat-intelligence reports.
The AI work stands on a data foundation. He led a team at the Digital Egypt Pioneers Initiative and won its Excellent Team Leader Award, then delivered more than twenty-five freelance projects with a 100% satisfaction record. He holds the IBM AI Engineering Professional Certificate alongside six IBM Generative AI certificates, and reads Mathematics & Computer Science at Helwan University.
The rooms are ordered the way the work was learned: Foundations first, where the gradient is computed by hand; then Retrieval, Perception, and Production, where those pieces are assembled into things other people can use. The Founder’s Room is last, because a company is what the rest of it becomes.
CURATORIAL TEXT · GALLERY WING · CAIRO
Five rooms. Select a room from the plan or the signage below; each wall is walked left to right.
ROOM I · FOUNDATIONS WING
Works whose entire point is the part that is usually imported: the autograd tape, the attention matrix, the sampler, the low-rank update. Nothing in this room is a wrapper.
A decoder-only transformer written from first principles. The tokenizer, the causal mask and the training loop are all the artist’s own hand.
Examine the workThe engine underneath everything else in this room. The whole autograd is verified against numerical gradients rather than trusted.
Examine the workImages generated from a prompt by walking a noise schedule backwards. The sampler is written out step by step, not called.
Examine the workSentiment and emotion classified with confidence that means something: the calibration curve is part of the work, not an afterthought.
Examine the workA LoRA fine-tuning, evaluation and serving pipeline. The low-rank update is derived and merged by hand; perplexity is computed, not reported.
Examine the workROOM II · RETRIEVAL WING
Index structures, ranking functions and the discipline of citation. Every work in this room can say where its answer came from.
Ask a PDF a question and get an answer with citations. The proximity graph that makes the search fast is built layer by layer.
Examine the workA retrieval system that plans, retrieves again, and then checks its own citations before it is allowed to answer.
Examine the workThe query is expanded through an entity graph before anything is retrieved, so neighbouring facts arrive with the obvious ones.
Examine the workSearch a codebase in plain language. Ranking is field-weighted, and the call graph is parsed from the syntax tree.
Examine the workLong text reduced to a summary and a list of action items. Sentence importance is an eigenvector, found by repeated multiplication.
Examine the workScores a résumé against a job description and names the gaps, matching skill to skill across a graph rather than by keyword.
Examine the workThe retrieval room’s working piece: a vector-search microservice with a REST surface, built to be deployed rather than demonstrated.
Examine the workROOM III · PERCEPTION WING
Vision and audio, approached from the signal upwards: filterbanks, region growing, connected components. Classical methods are not nostalgia here; they are the parts that can be explained.
Captions an image and answers questions about it, then shows you where it was looking. The saliency map is the honest part.
Examine the workSegmentation without a network: you click, and a minimum cut separates foreground from background across a pixel graph.
Examine the workIndustrial defect detection with a loop that asks a human about the cases it is least sure of, and learns from those first.
Examine the workAudio transcribed and then summarised. The front-end is built from the transform up: window, spectrum, mel bands, speech or silence.
Examine the workA voice assistant that can be interrupted. Barge-in is the hard requirement, and it is the state machine that makes it possible.
Examine the workROOM IV · PRODUCTION WING
Services with contracts: an endpoint, a schema, a statistic you can defend. Each of these has an evaluation harness or a statistics engine written alongside it.
Invoices and receipts turned into structured JSON with a confidence score per field, so a human knows what to check.
Examine the workA streaming moderation API that explains its decisions. The rule language and the classifier are both legible to a policy team.
Examine the workA recommendation service with its own offline evaluation harness — because a recommender without a metric is a guess.
Examine the workMulti-horizon forecasts served with prediction intervals and a backtest, so the uncertainty is on the record.
Examine the workMonitoring for models and the data feeding them: distribution shift measured, not assumed, and escalated by rule.
Examine the workAn A/B platform with a full statistics engine underneath it, including the variance reduction that makes small effects visible.
Examine the workDetects personal data, redacts it, and can put it back — with a risk score that says how re-identifiable the result still is.
Examine the workROOM V · FOUNDER’S ROOM
A company, an assistant and an agent. These are products rather than studies — architected end to end and hung together because that is what the rest of the collection becomes.
An AI cybersecurity platform, architected end to end. Real-time detection joins deep learning to behavioural NLP; an LLM layer writes the report.
Visit malwareguard.techAn assistant built only for data work: no general chat, just the analytical loop from raw table to cleaned frame to dashboard.
Examine the workMakes small local models call tools reliably by compiling the tool schema into a grammar, so invalid output cannot be sampled at all.
Examine the workSix bodies of technique held by the studio, with the state each is maintained in and the works that depend on it.
Transformers · autodiff · backpropagation · LoRA
Retrieval-augmented generation · fine-tuning · HuggingFace · LangChain
HNSW · BM25 · TextRank
Graph-cut · Grad-CAM · defect QA
Power BI · Tableau · SQL · Pandas
FastAPI · Docker · drift detection
Where the work was made, and in whose service. Listed most recent first.
2025 — present
MalwareGuard
Nov 2025 — Jul 2026
DEPI · Digital Egypt Pioneers Initiative
Oct 2024 — present
Mostaql & Khamsat
May 2024 — Sep 2024
DEPI · Digital Egypt Pioneers Initiative
Credentials on loan to this collection, with their lenders.
IBM · COURSERA · 2026
IBM · 2026
IBM · 2026
MAVEN ANALYTICS · 2024
IBM · FULL SERIES
CISCO · UNIVERSITY OF COLORADO BOULDER
Loans, commissions and appointments are handled directly by the maker.
Building something with AI or data? Let’s talk.
If you are hiring for AI/ML or data roles, or want a collaborator who builds the hard parts himself, get in touch. I read every message and reply.