MahmoudEl-Sherbiny

I build language models that understand Arabic and still fit on a small GPU.

Scroll. The points behind this page are a model’s embedding space.

Research to production

I’m a machine learning engineer working on LLMs, retrieval and reinforcement learning. I take a model from the paper stage to something people can open on their phone, and I make it small enough to run where compute is scarce.

Cairo-based, fluent in Arabic and English. B.S. in Computer Science, Egypt-Japan University of Science and Technology.

25%lower word error rate on Arabic speech
92%+accuracy reading prescriptions
75%less GPU memory for retrieval
3.72GPA out of 4.0

Where I’ve worked

MedDigi-Verse

Co-founder and AI engineer

  • Shipped a pharmacy assistant app on Google Play and the App Store. It reads prescriptions with 92%+ OCR accuracy and searches 25K+ products by meaning.
  • Coordinated a 20-person team across machine learning and full-stack development.

Edafa Telecom Solutions

Software engineering intern

  • Built a real-time health monitoring dashboard with Flutter, Spring Boot and PostgreSQL.
  • Designed a market simulation with five user roles, role-based access control and a scalable backend.

PSCwork

Client relationship manager

  • Ran onboarding and support for 200+ clients and improved workflows across teams.

Projects

Islam Web: Arabic question answering

2025

The problem: finding a relevant ruling among hundreds of thousands, in Arabic, on modest hardware.

  • Answers questions over 432K+ fatwas using hybrid retrieval: FAISS, BM25 and a reranker.
  • Fine-tuned Qwen2.5-1.5B on 10K QA samples, then pruned and quantized it.
Model memory6 GB before, 1.2 GB after
FAISSBM25RerankingQwen2.5-1.5BPruning

Audio Retrieval: Arabic speech pipeline

2025

The problem: noisy Arabic recordings that general speech models transcribe poorly and that are hard to search.

  • Tuned Whisper-Small (244M) with QLoRA plus denoising, voice activity detection and chunking. Word error rate dropped by about 25%.
  • Semantic search with FAISS and BGE-M3 embeddings in 4-bit, using about 75% less GPU memory.
Whisper-SmallQLoRAFAISSBGE-M34-bit

HackOps: AI cybersecurity platform

2025 – 2026

The problem: training an agent to plan attacks, then hosting it so teams can compete and be scored.

  • Reinforcement learning attack framework: Recurrent PPO paired with Qwen2.5-7B (LoRA, 4-bit). 10% better than baseline over 500K training steps.
  • Full-stack platform on Flask, React, MariaDB and Docker, with scoring, logging and a leaderboard.
Recurrent PPOQwen2.5-7BLoRAFlaskReactMariaDBDocker

JobFit: AI career copilot

2026

The problem: tailoring a resume to a job without inventing experience.

  • Job matching with Apify, multi-factor scoring and Gemini: 100% Precision@5 and 0.942 NDCG@5.
  • Resume tailoring that checks every claim against evidence. Across 225 evaluations, 100% of facts were preserved and none were unsupported.
ApifyGeminiMulti-factor scoringClaim verification

Results at a glance

ProjectMeasureResult
Audio RetrievalWord error rateabout 25% lower
Audio RetrievalGPU memoryabout 75% lower
Islam WebModel memory6 GB to 1.2 GB
MedDigi-VersePrescription OCR accuracy92%+
HackOpsVersus baseline, 500K steps10% better
JobFitPrecision@5 / NDCG@5100% / 0.942
JobFitFact preservation, 225 evaluations100%

What I work with

Machine learning

PyTorch, TensorFlow, Scikit-learn, HuggingFace, LangChain, FAISS. LLMs, RAG, reinforcement learning, NLP, generative AI, fine-tuning, quantization, pruning, ASR, OCR.

Engineering

Python, SQL, C++, Java, FastAPI, Flask, React, TypeScript, PostgreSQL, REST APIs, Docker, Git, Linux, Pandas, NumPy.

Languages

Arabic (native), English (fluent).

Education

Egypt-Japan University of Science and Technology

B.S. Computer Science, GPA 3.72/4.0. Class of June 2026.

Coursework: deep learning, NLP, reinforcement learning, computer vision, distributed systems, algorithms.

Beyond code

IEEE Computer Society, EJUST

Head of the AI and ML community. Led a 10-member team, mentored 100+ students and ran 6+ workshops.

Certified: Machine Learning Specialization (DeepLearning.AI and Harvard), DEPI Generative AI Track.

Have a model that needs to ship?

mahmoud.elsherbiny.work@gmail.com