REACT / PYTHON / APPLIED AIBASED IN CAIRO. OPEN TO RELOCATION.SCROLL TO MEET ME ↓
01 — THE PERSONA LITTLE CONTEXT
Hi, I’m Abdelrahman.
A product engineer who cares about the whole picture.
I build full-stack products and the systems behind them, from real-time applications and payment workflows to AI agents and event-driven platforms.
At Cravit, I’ve taken a restaurant platform from concept to production and led its backend re-platform. My work also spans AI infrastructure at Rafeeq and children’s recitation models at Ozzie.
02 — THE AI ENGINEERING STACKFROM INTERFACE TO INFERENCE
The full stack. The whole picture.
52 tools & practices. 13 connected disciplines.
An AI product spans the interface, services, data, models, evaluation, and operations. Explore my documented work alongside examples from the broader AI engineering ecosystem.
From my work reflects my documented experience. AI ecosystem introduces the wider landscape.
Tap any card for its role, practical context, and official documentation.
EXPLORE THE AI ENGINEERING TOOLKIT{ }
Explore every layer of an AI product, from interface to operations.
52 of 52 tools & practices
No tools match these filters.
Try another search or explore a different discipline.
Ecosystem entries are examples to explore. Documentation explains the tools; my résumé is the source for work experience.
03 — SELECTED WORKIDEAS, OUT IN THE WORLD
Things I’ve helped build.
Real products. Complex systems. Made for the people on the other side.
Selected work from my roles at Cravit, Rafeeq, and Ozzie. Visuals are illustrative.
04 — THE JOURNEYALWAYS BUILDING
Experience that adds up.
Product, platform, and applied AI. A growing perspective on engineering.
Aug 2026 – PresentRemote · Part-timeOzzieNLP EngineerEnd-to-end ownership of a children’s Quran recitation model, from data annotation to production.+
Own the recitation model end to end, from annotating children’s Quran recitation data through fine-tuning and production deployment.
The model has reached 85% accuracy on children’s recitations, with a target of 98%.
Jun 2025 – PresentCairo, EgyptCravitFounding EngineerFrom concept to production: a restaurant platform, an event-driven backend, and the AI systems behind it.+
Led the full re-platform from a NestJS monolith to an event-driven Python/FastAPI monorepo using async SQLAlchemy, Celery, Redis Pub/Sub, and PostgreSQL.
Platform scale: 2,100+ REST endpoints, 50+ business modules, 600+ ORM models, 350+ Alembic migrations, and 6,500+ automated tests.
Cut API response times 43%, average query execution 61%, and database CPU usage 38% while sustaining production availability.
Built a central AI gateway with model routing, fallbacks, circuit breakers, Redis LLM caching, 4 Qdrant vector collections, a Two-Tower recommendation service, and a dbt analytics warehouse.
Cut per-conversation inference cost by over 70% while holding answer quality flat through multi-tier routing, automatic fallback, prompt caching, tenant token budgets, response-format trimming, and cost-per-resolution instrumentation.
Shipped LangGraph marketing and operations agents with per-tenant budgets, human approval queues, guardrails, audit logging, fallbacks, circuit breakers, and an outage watchdog.
Integrated Paymob BNPL across eligibility, checkout, installment schedules, HMAC-authenticated webhooks, idempotency, reconciliation, refunds, and reporting; built restaurant customer subscriptions and restaurant SaaS subscriptions.
Owned concept-to-production delivery for a multi-tenant restaurant platform serving 14 restaurants, 22,000+ customer accounts, and 38,000+ orders; shipped 4 Next.js/React/TypeScript applications with real-time updates and 15+ integrations.
Fine-tuned transformer classifiers for Arabic, Egyptian dialect, Arabizi, and mixed Arabic/English intent, language, and urgency; added real-time voice with Deepgram STT and Azure TTS over WebSockets.
Jul 2026 – PresentRemote · ContractRafeeqProduct/GTM Engineer (Founding Team)AI product infrastructure, long-term user memory, and dependable delivery for a personal development platform.+
Architected the FastAPI backend for a personal development platform with 144 REST endpoints, 29 SQLAlchemy models, and 44 service modules.
Built and led the AI platform, integrating OpenAI, Anthropic, Gemini, and open-source models with routing, fallback, Qdrant retrieval, Redis session state, long-term memory, usage metering, and auditable tenant controls.
Built Docker-based exact-commit releases, health checks, authenticated backup, restore, and rollback workflows.
Preserved 4,573 data assertions through a production-copy CRM migration; covered the WhatsApp workspace with 703 Python and 71 TypeScript/Node unit tests.
Engineered the WhatsApp Business workspace and consent-first enrichment pipeline with suppression controls, encrypted contact data, PII isolation, recorded approvals, spend and usage ledgers, idempotency, reconciliation, and auditable fail-closed behavior.
Jun 2024 – Jun 2025RemoteStyle ProtocolFull Stack EngineerFull-stack product development and multi-chain smart contracts.+
Built a Next.js frontend and server APIs, increasing engagement 30%, reducing page-load time 20%, and cutting server response time 25%.
Deployed 6 smart contracts across Ethereum, Arbitrum, and Solana supporting 1,000+ transactions without operational issues.
Mar 2024 – Jun 2024RemoteNethermindBlockchain DeveloperSmart contract auditing, vulnerability remediation, and secure-coding workshops.+
Identified and helped remediate 2 high-severity DeFi vulnerabilities that could have caused multi-million-dollar losses.
Audited 7 Ethereum and Starknet projects and led secure-coding workshops.
Nov 2023 – Mar 2024RemoteRhodes AcademyWeb3 DeveloperDecentralized identity, tuition payments, and a virtual gallery.+
Shipped a decentralized identity system that reduced onboarding administration 30%, a DeFi tuition-payment flow, and a blockchain-based virtual gallery.
Feb 2023 – Oct 2023Cairo, EgyptEJADAFull Stack DeveloperReal-time learning experiences, faster data access, and containerized delivery.+
Shipped real-time chat for an e-learning platform, increasing engagement 20% and satisfaction 15%.
Improved query speed 30%, cut database use 40%, and containerized 4 services to reduce deployment time 50% and downtime 60%.
EDUCATION / 2019 — 2023
BSc, Computer Engineering
Alexandria University
CERTIFICATION
IBM AI Engineering
Professional Certificate · IBM / Coursera
05 — THE IMPACTTHE WORK ADDS UP
Good engineering. Measurable impact.
At Cravit: faster systems, lower costs, and confidence at scale.
43%
Lower API response time
Through an event-driven backend re-platform.
61%
Lower average query time
With 38% less database CPU use.
70%+
Lower AI inference cost
Per conversation, with answer quality held flat.
6,500+
Automated tests
Supporting the platform migration.
06 — YOUR NEXT IDEASTARTS WITH A HELLO
Let’s build something that matters.✳
Have a product, a platform, or a difficult engineering problem in mind? Let’s talk.
JavaScript server runtimes and structured backend modules.
Node.js supports server-side JavaScript, while NestJS organizes applications into modules and services. My resume lists both and documents Cravit's earlier NestJS monolith.
In my work
Node.js and NestJS in the resume; NestJS-to-Python re-platforming at Cravit.
Python database access and ORM models with async support.
Define persistence models and database operations alongside application services. My resume documents async SQLAlchemy at Cravit and 29 SQLAlchemy models at Rafeeq.
In my work
600+ ORM models at Cravit; 29 SQLAlchemy models at Rafeeq.
Fast caching, session state, and publish/subscribe messaging.
Reuse frequently needed results and coordinate realtime events through Pub/Sub. My resume documents LLM caching at Cravit and Redis session state at Rafeeq.
In my work
Redis LLM caching and Pub/Sub at Cravit; session state at Rafeeq.
Vector search for retrieval and AI application memory.
Search stored vectors and metadata for context relevant to a query. My resume documents four Qdrant collections at Cravit and Qdrant retrieval at Rafeeq.
In my work
Four Qdrant vector collections at Cravit; retrieval at Rafeeq.
Ground answers in retrieved context and retain useful conversation state.
Retrieve relevant information before generating a response and keep useful memory across interactions. Rafeeq's AI platform includes Qdrant retrieval, Redis session state, and long-term memory.
In my work
RAG in the resume skill list; retrieval and long-term memory at Rafeeq.
An open-source framework for RAG pipelines and data-connected agents.
Connect ingestion, indexing, retrieval, and generation into an application workflow. This is an ecosystem example for building around private or domain-specific information.
Reorder retrieved documents by semantic relevance to a query.
Score candidate documents after initial retrieval to surface more relevant context. This is an ecosystem example of the reranking stage in a RAG pipeline.
Stateful agent workflows with durable execution and human review.
Coordinate agent steps, state, and interruptions for review. My work at Cravit includes LangGraph marketing and operations agents with production controls.
In my work
LangGraph agents at Cravit with approval queues, budgets, fallbacks, and audit logging.
Claude model APIs for language and agent applications.
Connect a model provider to application workflows and routing policies. My resume documents Anthropic integration alongside other model providers at Rafeeq.
In my work
Anthropic integration at Rafeeq; Claude in the resume skill list.
Composable language-model programs with metric-driven optimization.
Describe tasks as modules and optimize program behavior against a chosen metric. This is an ecosystem example for structured prompting and systematic improvement.
Persistent two-way communication for live interactions.
Exchange events and streaming data without creating a new connection for each message. My work includes real-time voice over WebSockets and real-time product updates.
In my work
Voice over WebSockets at Cravit; realtime application and chat work in the resume.
Intent, language, and urgency classification across mixed-language input.
Adapt transformer classifiers to Arabic, Egyptian dialect, Arabizi, and mixed Arabic/English text. The documented tasks at Cravit cover intent, language, and urgency.
In my work
Transformer classifiers for Arabic, Egyptian dialect, Arabizi, and mixed Arabic/English at Cravit.
Pretrained model definitions and utilities for inference and training.
Work with model architectures across text, audio, vision, and multimodal tasks. This library is an ecosystem example; the resume documents transformer work without naming the library.
Text-processing pipelines for tokens, entities, and linguistic features.
Process documents with components such as tokenization, tagging, and named-entity recognition. This is an ecosystem example for conventional NLP alongside LLM applications.
Adapt pretrained models to task-specific language and speech data.
My documented work includes fine-tuned multilingual classifiers and end-to-end recitation modeling. Ozzie's children's Quran recitation model has reached 85% accuracy.
In my work
Fine-tuning at Cravit and Ozzie; Ozzie's achieved accuracy is 85%.
Tensor operations, automatic differentiation, and neural-network training.
Build training and inference workflows around tensors and neural-network modules. This is an ecosystem example for the training layer of AI engineering.
Efficient adaptation using a smaller set of trainable parameters.
Adapt a pretrained model through methods such as low-rank adapters. This is an ecosystem example that can reduce training and artifact storage requirements.
Post-training tools for supervised and preference-based model adaptation.
Use trainers for supervised fine-tuning and methods such as direct preference optimization. This is an ecosystem example for the model post-training stage.
Quality checks that make model and prompt changes measurable.
Evaluate changes against representative tasks before releasing them. Evaluation gates appear in my resume, alongside work that held answer quality flat while reducing inference cost.
In my work
Evaluation gates in the resume skill list; answer quality held flat during Cravit's cost reduction.
Regression tests for API, data, and product behavior.
Test deterministic application behavior alongside model evaluation. My resume documents more than 6,500 automated tests at Cravit and 4,573 preserved data assertions in Rafeeq's migration.
In my work
6,500+ automated tests at Cravit; 703 Python and 71 TypeScript/Node unit tests at Rafeeq.
Evaluation metrics and experiments for RAG and agent applications.
Measure dimensions such as context precision, recall, and answer faithfulness. This is an ecosystem example for comparing retrieval and generation changes.
Compare prompts and models through evaluations and red-team tests.
Run test cases across prompt and provider configurations, then compare results. This is an ecosystem example for regression checks and adversarial testing.
Experiment tracking, model packaging, and lifecycle management.
Track parameters and results while managing model artifacts and registry workflows. This is an ecosystem example for making model development reproducible.
Version data, models, and reproducible machine-learning pipelines.
Track data and model assets through metadata alongside source control. This is an ecosystem example for connecting a trained model to its data and pipeline version.
LLM inference serving with batching and efficient memory management.
Serve language models with continuous batching, streaming, and prefix caching. This is an ecosystem example for improving inference throughput in a self-hosted stack.
Containerized services and reproducible application releases.
Package application code and dependencies into deployable containers. My work at Rafeeq includes Docker-based exact-commit releases with health checks, backups, restore, and rollback.
In my work
Docker releases at Rafeeq; containerized services at EJADA.
Repeatable delivery with health checks, backups, and rollback.
Automate checks and make release recovery part of the delivery workflow. CI/CD appears in my resume, with concrete health-check and recovery work at Rafeeq.
In my work
CI/CD in resume skills; health checks, authenticated backups, restore, and rollback at Rafeeq.
Infrastructure as code for versioned cloud resources.
Define infrastructure in configuration and review planned changes before applying them. This is an ecosystem example for repeatable environments and infrastructure delivery.
Review checkpoints for agent actions that need human judgment.
Pause a workflow when an action requires approval and resume after a decision. My resume documents human approval queues in Cravit agents and recorded approvals in Rafeeq workflows.
In my work
Approval queues at Cravit; recorded approvals and consent-first workflows at Rafeeq.
Isolation and privacy controls for multi-tenant AI products.
Keep tenant data and sensitive contact information scoped to the appropriate workflows. My resume documents encrypted contact data, PII isolation, suppression controls, and fail-closed behavior at Rafeeq.
In my work
Auditable tenant controls, encrypted contact data, and PII isolation at Rafeeq.
Validate agent behavior and retain records of important actions.
Combine workflow checks with audit logging to make operations inspectable. Cravit's documented agents include guardrails, audit logs, circuit breakers, and an outage watchdog.
In my work
Guardrails and audit logging in Cravit's production agents.
Route requests across model tiers with resilient fallback behavior.
Cravit's gateway combined multi-tier routing, automatic fallbacks, and circuit breakers. With caching, budgets, response trimming, and instrumentation, this cut per-conversation inference cost by over 70% while holding answer quality flat.
In my work
Over 70% lower per-conversation inference cost at Cravit, with answer quality held flat.
Track token use, tenant spend, and cost per resolved outcome.
Apply tenant token budgets and trim response formats to control resource use. My resume documents cost-per-resolution instrumentation at Cravit and spend and usage ledgers at Rafeeq.
In my work
Tenant budgets and cost-per-resolution at Cravit; spend and usage ledgers at Rafeeq.
Use lower-precision model representations to reduce memory requirements.
Choose a quantization method according to hardware and model constraints. This is an ecosystem example whose quality tradeoffs need evaluation on the target task.
Concept-to-production engineering for a multi-tenant restaurant platform.
14restaurants
22K+customer accounts
38K+orders
What I built
Four Next.js, React, and TypeScript applications with real-time updates and more than 15 external integrations. Payment workflows include Paymob BNPL, restaurant subscriptions, authenticated webhooks, idempotency, reconciliation, and refunds.
The platform underneath
Led the move from a NestJS monolith to an event-driven Python/FastAPI monorepo with async SQLAlchemy, Celery, Redis Pub/Sub, and PostgreSQL. The platform spans 2,100+ REST endpoints and 50+ business modules, backed by 6,500+ automated tests.
43% lower API response time · 61% lower average query time · 38% less database CPU use
Selected work from my Founding Engineer role at Cravit. Visuals on this portfolio are illustrative, rather than product screenshots.
SELECTED WORK / CRAVIT
Smarter AI. Better economics.
A central AI gateway and agents designed around cost, quality, and reliability.
70%+lower inference cost per conversation
4Qdrant vector collections
How it works
Multi-tier model routing with automatic fallback, Redis LLM caching, prompt caching against byte-stable system prefixes, per-tenant token budgets, and response-format trimming.
Cost per resolution is instrumented alongside latency and quality. Answer quality was held flat while per-conversation inference cost fell by over 70%.
Beyond the gateway
LangGraph marketing and operations agents use human approval queues, guardrails, audit logging, circuit breakers, and an outage watchdog. The wider platform includes a Two-Tower recommendation service and a dbt analytics warehouse.
Selected work from my Founding Engineer role at Cravit. The architecture illustration summarizes documented components.
SELECTED WORK / RAFEEQ
AI with context.
The backend and AI platform for a personal development product.
144REST endpoints
29data models
44service modules
Product infrastructure
Multi-provider routing across OpenAI, Anthropic, Gemini, and open-source models; automatic fallback; Qdrant retrieval; Redis session state; and long-term user memory. Usage metering and auditable tenant controls support the platform.
Dependable delivery
Docker-based exact-commit releases with health checks, authenticated backup, restore, and rollback. A production-copy CRM migration preserved 4,573 data assertions, and the WhatsApp workspace has 703 Python and 71 TypeScript/Node unit tests.
Selected work from my Product/GTM Engineer role on Rafeeq’s founding team.
SELECTED WORK / OZZIE
Learning, heard.
An NLP model for children’s Quran recitation.
85%documented accuracy on children’s recitations
End-to-end ownership
My work spans annotating recitation data, fine-tuning the model, and deploying it to production. The model has reached 85% accuracy on children’s recitations; improving that performance is ongoing.
Selected work from my part-time NLP Engineer role at Ozzie. The waveform is an illustrative graphic.