Ten years, in order

Founding AI Engineer (Contract), Stealth Startup

• Built and deployed LLM systems for 14+ isolated enterprise customers.
• Co-developed a LangChain Deep Agents multi-agent system with memory management, covering structured and unstructured retrieval, SaaS integrations, and report generation.
• Built an integrations agent with read and write access across connected SaaS systems — Google Workspace, Slack, Notion, GitHub, and Linear.
• Delivered the Composio-based integration layer behind it, with per-user RBAC, human approval gates, and white-labelled OAuth.
• Raised first-pass document generation success to near 100% through per-format validation loops, visual inspection of rendered output, and self-repair.
• Built an automated, fault-tolerant enterprise data onboarding pipeline — schema discovery, embeddings, metadata preparation, and relationship inference — cutting onboarding from hours to minutes.
• Improved retrieval quality with hybrid search, re-ranking, and filesystem-structure facets for path-based scoping, reaching near-complete recall on customer-defined evaluation queries.

• LangChain
• LangGraph
• Claude Agent SDK
• RAG
• Composio

Senior Data Scientist (NLP), Implicit(formerly Agolo)

• Built Graph RAG and question answering for Implicit’s Answers product from scratch on LangGraph and Neo4j.
• Raised QA correctness (LLM-judged) by more than 50% on the internal evaluation set by tuning graph retrieval and agent behaviour: grounding answers in customer context, raising clarifying questions only on genuine ambiguity, suppressing irrelevant queries, and enforcing guardrails.
• Deployed and served the QA stack on a self-hosted Llama model using vLLM, and later on Amazon Bedrock.
• Co-authored FactAlign, a knowledge-graph-alignment method for fact-level hallucination detection, reaching 0.889 F1 on detection and 0.825 F1 on intrinsic/extrinsic classification with no fine-tuning and no repeated sampling.
• Fine-tuned LLMs on KG-to-text for entity-centric summarization, surpassing the then state of the art. Read the write-up ↗
• Introduced LLMs across summarization, knowledge graph construction, and question answering; maintained and extended the extractive summarization platform for new customer use cases.

• Knowledge Graphs
• Neo4j
• Graph RAG
• LangGraph
• vLLM
• Amazon Bedrock

• Led toFactAlign: Fact-Level Hallucination Detection and Classification Through Knowledge Graph Alignment · TrustNLP 2024

Data Scientist, DataPlus

• Built ETL and predictive-analytics workflows and dashboards for financial, telecom, and government clients.

• Tableau
• Alteryx

Google Summer of Code Student, Distributed Red Hen Lab

• Implemented an Arabic speech recognition system for broadcast speech using Kaldi and VariKN.
• Implemented an Arabic dialect identification system using Kaldi, Scikit-learn, and Keras.
• Mentored several projects in subsequent GSoC programs.

• Kaldi
• VariKN
• Scikit-learn
• Keras

Research Assistant, Cairo University, Faculty of Engineering

• Co-developed Senteech, a speech emotion classification system for customer-service calls in low-resource languages.

• Scikit-learn
• TensorFlow
• openSMILE

Teaching Assistant, Cairo University, Faculty of Engineering

• Taught CMP302 — Advanced Algorithms.

Machine Learning Researcher, RDI Egypt

• Raised mispronunciation detection accuracy by ~2% for Hafss, a Holy Quran recitation training system, by training Kaldi acoustic models on then state-of-the-art architectures and expanding the training data. Hafss ↗
• Maintained and extended Hafss’s C++ language model toolkit, fixing and expanding its generation rules.
• Built a .NET annotation tool for the linguistics team.

• Kaldi
• C++
• .NET