Get in touch
Full-Stack Engineer | RAG & Agentic AI Systems

Abdelrahman
Yasser Mostafa

Full-stack engineer who builds end-to-end
from custom-trained language models, to production RAG infrastructure, to the products built on top of them. Comfortable owning a project across model training, backend, and frontend.

LinkedIn GitHub
Giza, Egypt +20 112 725 6282 Hugging Face
abddo.py 
class Engineer:
    stack = ["Python", "TypeScript", "Rust"]
    focus = ["Ai integrated smart systems", "RAG", "Product"]
    location = "Giza, EG"
   
    def build(self, end_to_end=True):
        # model training → backend → frontend
        return self.ship()

abddo = Engineer()
abddo.build()

EXPERIENCE

Building products from model to interface.

01

Full-Stack Engineer

Bar3

  • Designed and built an Arabic-first, personalized exam platform for Egyptian high-school students — custom exams and diagnostic reports targeting each student's weak points.
  • Implemented an agentic RAG-powered chat with dynamic skill injection per question type and visible step-by-step planning in the UI.
  • Built the full stack on top of Rx (retrieval infrastructure) and Space-v1 (embeddings) — one integrated AI product from model to interface.
02

Full-Stack Engineer

Rx — Retrieval x

  • Built a multi-tenant RAG-as-a-Service platform from scratch: add data, get an API key, and have a production RAG pipeline running in under 10 lines of SDK code (Python & TypeScript).
  • Implemented hybrid BM25 + vector search, per-tenant user management with role-based permissions, billing, and audit logging.
  • Serves as the retrieval backbone for EDUx; stress-tested across additional verticals — real estate, HR, customer support, education, and more.

SKILLS

End-to-end ownership.

Languages

Python, TypeScript / JavaScript, Rust

Backend

Multi-tenancy + RBAC, FastAPI + Pydantic, PostgreSQL + pgvector, MinIO/R2, Redis, billing & audit logging, API/SDK design, clean architecture, Nginx + Cloudflare.

AI / ML

RAG & hybrid search (BM25 + vector), agentic/LLM planning loops (Gemini API), embeddings, model distillation, LoRA fine-tuning, STT/TTS voice pipelines.

Frontend & Tooling

HTML, CSS, JavaScript, Next.js, React, shadcn. Playwright, tree-sitter / AST analysis, Docker, Git. Deployed on AWS, Vercel, Hostinger.

stack.toml config 
[languages]
primary = ["Python", "TypeScript", "Rust"]

[backend]
framework = "FastAPI"
database = "PostgreSQL + pgvector"
storage = "MinIO / R2"
cache = "Redis"

[ai]
search = "hybrid BM25 + vector"
models = ["MathCore", "Space-v1"]
training = ["LoRA", "distillation"]
tools = ["Claude code", "codex", "Cursor", "Claude code"]
used = ["Claude", "ChatGPT", "Qwen", "DeepSeek" , "GLM", "Kimi"] 
      # im actually trying any new model that comes out
[frontend]
framework = "Next.js + React"
ui = "shadcn"

[deploy]
targets = ["AWS", "Vercel", "Hostinger"]

PROJECTS

Additional work beyond the day job.

01

Space-v1

~80M-param Recurrent-Depth Transformer distilled from Qwen-4B, built with Claude (Fable 5). Runs efficiently on CPU and powers real-time retrieval inside EDUx, matching much larger models on complex queries.

02

MathCore

Neural arithmetic-reasoning model testing generalization vs. memorization: 99.3598% accuracy (6402/1M) on unseen addition/subtraction problems. Built with Claude (Fable 5).

03

mcp-checker

Multi-language security scanner for MCP servers: AST-based static analysis (tree-sitter) across Python/JS/TS/Rust/Go, plus an optional AI-powered taint-analysis layer.

04

browser-agent

Agentic browser-automation tool built on Playwright; an LLM planning loop observes each page and decides the next action step by step until the task completes.

05

AI Customer Service Pipeline

Voice-based support system with a full STT → RAG/LLM → TTS pipeline and live chat.

06

rusty-excel

Web app to upload a spreadsheet, select columns, and search by key via hash-map lookups — ~9ms response time. Next.js + Rust backend.

EDUCATION

Information Systems, BIS

In progress — 3rd year