AI & Backend Engineer
aka Romia
I build LLM agents, RAG systems, and the backends they run on. At Zaylon AI, the company I co-founded, the agent I built takes a shopper from the first product question to a paid order inside WhatsApp and Instagram, in their own dialect.
Projects
9 projects
Zaylon AI
An AI that sells over WhatsApp and Instagram for MENA merchants, carrying a customer from the first product question to a confirmed payment, all inside the chat.
LangGraph tool-calling agent and tri-lingual NLP on a multi-tenant FastAPI backend
Read the write-upcommentdraft
A command line tool that triages comments on your own posts into reply, skip, or escalate, drafts the reply from one source document, and sends nothing a person has not approved one keystroke at a time.
Eight platform connection guides written from primary sources, and a publish path with no way to approve in bulk: one keystroke, one reply
Read the write-upGraduation project (team): a team chat platform with a workspace-grounded RAG assistant that answers from your own documents, with citations. I owned the AI, retrieval, evaluation, and deployment.
Found the root cause of weak answers (over-fragmented chunks) and proved the fix with a paired, Holm-corrected eval: judged correctness 0.657 to 0.855 on the workspace's own corpus.
Read the write-upContextIQ: Hybrid-Retrieval RAG
A worked example of production RAG that runs on CPU: dense and lexical retrieval fused with reciprocal rank fusion, cross-encoder reranking, grounded citations, and an evaluation harness that measures the difference.
Hybrid retrieval plus reranking measured on a confusable corpus: best precision at the top, hit@3 0.83 and MRR 0.78, against a dense-only baseline at 0.67
Virtual Banking Microservices
An event-driven banking system: a Spring Cloud Gateway fronting four core services, with a transfer saga over Kafka that holds up under concurrency, retries, and restarts, and tests that prove it.
A transfer saga with a transactional outbox, idempotent consumers, and pessimistic locking; a Testcontainers test fires twenty simultaneous transfers and proves no double-spend.
Read the write-upMultimodal Emotion Recognition
Speaker-independent speech emotion recognition on RAVDESS, rebuilt to remove a speaker-leakage flaw that inflates most published results. WavLM-large audio with learnable layer-weighting, fused with a facial-expression model.
Honest speaker-independent: 78.8% audio-visual (calibrated late fusion) on a 70.3% audio base, all leak-free, where the common random split would fake the audio up to ~78% by memorizing speakers
VoicePrint: Authorship Stylometry
An engine that measures the fingerprint of how you write from your own samples, then adapts your own drafts toward that voice. Interpretable stylometric features paired with a StyleDistance neural style embedding and an in-context rewrite.
With zero content words, a 130-dimensional function-word vector separated five authors at 0.684 macro-F1 and 0.889 accuracy, against a 0.20 five-class baseline
Games and puzzles
PlayableTwo classical search problems, both running entirely in your browser.
CSP Sudoku Solver
Constraint satisfaction solver using backtracking, AC-3 arc consistency, and MRV heuristic for efficient puzzle solving.
Zaylon AI
Conversational commerce for MENA merchants, on WhatsApp and Instagram
Zaylon turns WhatsApp and Instagram into full sales channels for MENA merchants. Customers browse, ask in their own dialect, and check out without ever leaving the chat. It reads Egyptian Arabic and Franco-Arabic the way a local would, matches a product from a photo, recovers carts on its own, and carries someone from the first question to a paid order. We built the dialect engine and the conversational commerce layer in-house.
0
Platforms
0
Channels
0
Payments
0
Dialects
What it does
Handles the whole conversation end to end: answers product questions, builds the cart, and closes the order, all inside the same chat and without a human stepping in.
DialectBridge detects whether a customer is writing English, Egyptian Arabic, or Franco-Arabic and answers back in the same dialect.
Customers can send a photo of a product, and Zaylon matches it against the merchant's catalog.
Follows up on abandoned carts and steps in when a conversation starts to turn negative.
A merchant dashboard with a live inbox, analytics, RFM segmentation, and A/B testing.
Each merchant's data is encrypted on its own and isolated at the row level, so tenants never see each other.
Integrations built
Built to plug into the MENA commerce stack
Messaging
E-Commerce
Payments
Shipping
Built with
Experience
Software Engineer & Co-Founder
Co-founded Zaylon and built its AI system and backend: a LangGraph tool-calling agent and tri-lingual NLP on a FastAPI backend with PostgreSQL/Redis, multi-tenant isolation, and Docker deployment, built for MENA merchants selling on WhatsApp, Instagram, and web.
- ▸Built the conversational AI in LangGraph as a single tool-calling agent with about 30 scoped tools across sales, support, and checkout, consolidated from an earlier multi-agent supervisor for reliability and easier reasoning
- ▸Built hybrid product search combining keyword matching, semantic vector similarity, and Reciprocal Rank Fusion scoring
- ▸Wrote the FastAPI Python backend and its service architecture, containerized with Docker for deployment
- ▸Built multi-tenant security: argon2id key auth, per-merchant row-level isolation across roughly 80 tables, and semgrep CI gates enforcing tenant scoping
- ▸Built a Postgres-backed saga engine with reverse-order compensation so a failed checkout step rolls back cleanly
- ▸Batched rapid-fire WhatsApp messages in Redis into a single turn before invoking the agent, so it reads the whole thought at once instead of replying to each fragment
AI Coding Expert
Scale AI · Contract
Reviewed and improved LLM-generated code for RLHF training data used to train frontier AI models.
- ▸Reviewed and ranked code repositories across languages for RLHF data pipelines used to train frontier AI models
- ▸Built test suites as quality control, scoring correctness, efficiency, style adherence, and edge case handling
- ▸Reviewed thousands of code samples, providing detailed feedback on reasoning chains, code structure, and technical accuracy
Software Integration Engineering Intern
Ejada
Worked on core banking integration systems, building APIs and event-driven messaging pipelines.
- ▸Built RESTful API endpoints for core banking operations using Spring Boot with query optimization and Redis caching for frequently accessed account data
- ▸Configured Apache Kafka producers and consumers for asynchronous transaction processing, decoupling payment events from the main request lifecycle
- ▸Implemented API gateway security with OAuth2 authentication, rate limiting, and request validation using WSO2 API Manager
Automotive Cybersecurity Intern
Swift-ACT
Contributed to automotive cybersecurity research, building security scenario catalogs and running attack simulations in vehicle simulation environments.
- ▸Built security scenario catalogs with threat models, attack vectors, and mitigations for CAN bus and ECU communication systems following automotive cybersecurity standards
- ▸Conducted fuzz testing on Automotive Emergency Braking (AEB) components, identifying edge cases in sensor fusion timing and brake actuation logic under corrupted input data
Tech Stack
Backend
Powers Zaylon AI, Virtual Banking, ContextIQ RAG
AI / ML
Powers Zaylon AI, ContextIQ RAG, Emotion Recognition
DevOps & Cloud
CI/CD and deployment across all projects
Languages
Primary: Python, Java, TypeScript. Growing: Go
About Romia
I co-founded Zaylon AI and built its AI system and backend: the LangGraph agent that sells over WhatsApp and Instagram, the tri-lingual NLP, and the FastAPI backend behind them. I work across the stack but lean toward backend and AI. B.Sc. in Computer and Communications Engineering with an AI concentration from Alexandria University. I speak English, Arabic, German, and Spanish, and I play guitar.
Education
B.Sc. in Computer and Communications Engineering, AI concentration
Alexandria University, Faculty of Engineering
GPA: 3.75 / 4.0
Spoken Languages
Certifications
Meta Back-end Developer Specialization
Coursera · 2023
Competitions & Hackathons
References
Dr. Mohamed Moustafa Mahmoud (Dr. 3M)
Global Performance Management Academy (GPMA)
Ph.D., University of Pennsylvania
Client. AI agents and automation for his organization.
Let's Connect
I like building systems that actually ship. If you're working on agents, backend infrastructure, or AI products, or hiring for one, say hi.
commentdraft
A command line tool that drafts replies to comments, and refuses to send them
It reads the comments on a creator's own posts and decides, for each one, whether to reply, skip, or escalate to a person. Every decision comes back with a one-line reason the operator can disagree with, and every draft is written from one source document they supply. What the tool is, more than any of that, is a set of refusals.
It will not send without you
No --yes, no --all, and no config key that changes it. Publishing thirty replies costs thirty keystrokes, which is the design and not an oversight. The release on PyPI today is the drafting half: it reads a CSV, writes a page of drafts, and exits.
It will not answer past the source
A draft may state only what your document states. A question the document does not cover becomes a person's to answer, never a guess.
It will not claim reach it lacks
One connector exists, for Facebook Pages, and it has never been run against a live Page. The other seven platforms have guides, not code.
One published run/11 comments
reply
“Does it cover mushrooms at all?”
content question about a subject the book deliberately excludes
skip
“This is a scam you are just here to sell a book”
accusation rather than an objection, answering it widens the argument
escalate
“I have a nut allergy is any of this dangerous for me”
health question beyond the scope of the book, directed to medical professional
Seven of the eleven got no draft at all.
Verbatim from a run of the shipped example, written up with the command that reproduces it.