Mohamed Rayen Fekih

Senior AI / Machine Learning Engineer

AI Engineer with 5+ years of experience building production machine learning and GenAI systems. My work spans semantic search, RAG applications, agentic workflows, computer vision, recommendation engines, forecasting, and large-scale content processing systems.

Career

Work Experience

AI Engineer

iHorizons

Jul 2025 - Present Qatar · Remote

Developed Arabic retrieval, content-processing, RAG, multi-agent, and computer vision systems for government and industry-focused use cases.

  • Deliver government-sector semantic search, RAG, and multi-agent applications.
  • Build content enrichment, classification, and automation pipelines with Google Cloud and Vertex AI.
  • Implement evaluation and human-review workflows for sensitive public-service use cases.
  • Develop and evaluate computer vision models, tracking workflows, and safety-event logic for recorded industrial footage.

Projects

Fatwa Semantic Search

Hybrid retrieval system for finding the most relevant fatwa across the Ministry's IslamWeb corpus. Visit FatwaTok

Client: Qatar Ministry of Awqaf and Islamic Affairs

Contribution: Built Gemini embedding retrieval, BM25 keyword retrieval, RRF fusion, retrieval evaluation workflows, human labeling support, LLM-assisted relevance review, and production search deployment.

PythonGemini EmbeddingsBigQuery Vector SearchBM25Hybrid RetrievalRRFGoogle Cloud
View project details
165K+ searchable Arabic fatwas
12s → 4s end-to-end response latency (67% reduction)
3 production FastAPI services: search, summarization, and Arabic TTS
3s hybrid search response time
10M+ average monthly visits to IslamWeb

Fatwa Content Processing Pipeline

Scheduled production pipeline for fetching new fatwas, enriching the content, generating summaries and audio, and refreshing search artifacts.

Client: Qatar Ministry of Awqaf and Islamic Affairs

Contribution: Built a daily Cloud Run pipeline that detects newly published fatwas, fetches and processes content, runs Gemini batch summarization, flags sensitive and modern topics, generates Arabic TTS with Quran-aware preprocessing, and refreshes BM25 search artifacts.

PythonGemini Batch InferenceGoogle Cloud SchedulerCloud RunCloud StorageBigQueryText-to-SpeechBM25
View project details
165K+ Fatwa text processed
Daily scheduled content synchronization
Arabic TTS + Quran-aware preprocessing

Oil Rig Safety Computer Vision

Computer vision safety POC that detects workers and PPE in recorded oil-rig CCTV footage, tracks workers across frames, and raises alerts when they enter predefined danger zones.

Contribution: Researched and prepared domain data, annotated and versioned datasets in Roboflow, fine-tuned YOLO and RF-DETR detectors, tracked evaluation with MLflow, and integrated the selected RF-DETR model with ByteTrack, zone-event logic, and OpenCV video rendering.

PythonPyTorchRF-DETRYOLOOpenCVByteTrackRoboflowMLflowObject DetectionMulti-Object Tracking
View project details
≈0.80 best RF-DETR EMA mAP on project validation data
3 detected classes: worker, hardhat, and vest
2 detector families fine-tuned and benchmarked

Football Match Analytics Computer Vision

Recorded-match computer vision pipeline that detects and tracks match entities, maps play to pitch coordinates, and renders numerical and spatial analytics.

Contribution: Fine-tuned YOLO 26 Large for four-class detection and built the two-pass pipeline covering ByteTrack continuity, bounded ball interpolation, appearance-based team assignment, homography, possession and movement estimates, tactical radar views, and annotated video output.

PythonPyTorchYOLO 26OpenCVByteTrackSupervisionSigLIPUMAPK-MeansScikit-learnHomographyRoboflowPillow
View project details
4 detected entity classes
3 core KPIs: possession, speed, and distance
2 tactical radar modes: Voronoi and heatmap

Khateeb Multi-Agent Assistant

A grounded Arabic sermon-generation assistant for Khateebs, using retrieval, specialist agents, moderation, and review loops to support Friday sermon preparation while preserving human accountability.

Client: Qatar Ministry of Awqaf and Islamic Affairs

Contribution: Built the n8n POC and helped move the production workflow to Google ADK and Agent Engine, separating requirement extraction, retrieval, drafting, moderation, revision, word-count adjustment, and human review.

Google ADKAgent EngineVertex AIGoogle RAG EngineGeminin8nArabic NLPMulti-Agent Workflows
View project details
1,000+ mosques in the intended Qatar support scope
4 collaborating specialist agents

IslamWeb RAG Chatbot

Arabic assistant grounded in an established corpus of religious rulings, with scholar escalation for sensitive requests.

Client: Qatar Ministry of Awqaf and Islamic Affairs

Contribution: Developed retrieval and response workflows, plus an intent agent that escalates appropriate conversations to a scholar.

PythonGemini EmbeddingsBigQuery Vector SearchTelegram
View project details

Machine Learning Engineer

Infor

May 2023 - Jun 2025 New York · Remote

Developed enterprise GenAI, pricing, and product recommendation systems.

  • Built and evaluated machine learning solutions for enterprise customers.
  • Optimized production inference and created automated delivery pipelines.
  • Worked across data analysis, modeling, deployment, and customer requirements; guided junior engineers.

Projects

GenAI Assistant Agent

Graph-based assistant with function calling, dynamic routing, and contextual memory.

Contribution: Developed the LangGraph workflow, memory retrieval, tool integration, and Docker deployment.

PythonLangchainLangGraph StudioLangChainDocker

Pricing System

Machine learning engine that recommends prices while balancing revenue and customer retention.

Contribution: Built product and customer segmentation models, evaluated performance, and deployed customer-specific solutions.

PythonScikit-learnPandasTableauClustering
3 Customer deployments
5% profit increase contributed to for an enterprise customer

Product Recommendation Systems

Upsell and cross-sell recommendation systems that increase basket size and revenue.

Contribution: Built recommendation models for upselling and cross-selling, optimized inference with Cython and NumPy to reduce response time by 80% and created CI/CD workflows.

PythonCythonNumPyCI/CD
80% reduction in response time

Data Scientist

Cognira

Feb 2021 - May 2023 Tunisia · On-site

Developed retail forecasting and manufacturing process-optimization solutions.

  • Translated customer requirements into applied machine learning solutions.
  • Prepared large operational datasets with PySpark and SQL.
  • Deployed models and automated recurring analytics workflows.
  • Monitored production models for performance degradation and training-serving skew using metric thresholds and automated alerts.

Projects

Retail Demand Forecasting

Scalable demand forecasting and promotion planning engine for large-scale retail operations.

Contribution: Contributed within a 7-person engineering team to scalable forecasting pipelines for product demand and promotional impact, working with clients to validate data and forecasts and building Plotly Dash and Tableau dashboards.

PythonPySparkSQLTime Series
10M+ transactions processed
3 Customer deployments

Process Quality Optimization

Regression and explainability system for improving cheese-manufacturing efficiency.

Contribution: Built the modeling and explainability modules and automated daily operational reporting.

PythonRegressionML ExplainabilityDashboards
15% profit-margin improvement
View experience page

Public libraries

Open Source

Reusable tools I build and publish for Arabic text processing and AI workflows.

hadith-detector

Offline Python library for detecting Arabic Hadith passages and retrieving their supporting sources, with exact and configurable fuzzy matching.

Contribution: Built and published v0.1.1 on PyPI, including Arabic normalization, positional indexing, ordered matching, source retrieval, original-text offsets, and HTML highlighting.

GenAI & NLP
PythonArabic NLPMARISARapidFuzz
14,940Hadith records from Bukhari and Muslim
1.04Mindexed Arabic tokens

LLM Fine-Tuning Pipelines

Personal research project covering GRPO fine-tuning on GSM8K and an SFT dataset pipeline for certification-question generation.

Contribution: Built a single-GPU Qwen2.5-3B-Instruct GRPO pipeline using Unsloth, TRL, QLoRA, vLLM generation, and five rule-based rewards, following the standard Unsloth recipe. Built resumable Gemini-based metadata extraction with Pydantic validation for SFT data. No training results or evaluation benchmarks are published.

GenAI & NLP
PythonUnslothTRLGRPOQLoRAvLLMGeminiPydantic
5rule-based reward functions
1 GPUtraining configuration

Technical profile

Skills

AI and Machine Learning

GenAIRAGSemantic SearchAgent WorkflowsComputer VisionObject DetectionMulti-Object TrackingRecommendation SystemsForecastingPricing ModelsScikit-learnGRPOQLoRALLM Evaluation

Programming and Data

PythonSQLPySparkPandasNumPyCython

Cloud and Deployment

Google CloudVertex AIBigQueryCloud RunCloud StorageDockerCI/CDGit

Frameworks and Tools

PyTorchOpenCVMLflowLangChainLangGraphn8nGoogle RAG EngineTableauGitLab

Background

Education

Bachelor’s Degree in Computer Science, Artificial Intelligence Specialization

National Engineering School of Tunis (ENIT)

2016 - 2021 Tunis, Tunisia

Certifications

Google Cloud Certified: Professional Machine Learning Engineer

Google Cloud · September 2026 – September 2028

View credential

Claude Certified Architect, Foundations (CCAR-F)

Anthropic · August 2026 – August 2027

View credential

Contact

Get in touch

Open to relocation. For job opportunities or professional conversations, contact me by email or LinkedIn.