[nom]

AI Engineer — Fintech AI & Applied GenAI

[téléphone] · [e-mail]

[lien] · [lien]

PROFESSIONAL SUMMARY

AI Engineer with 2+ years building production AI systems in fintech, specializing in computer vision, document AI, and LLM-powered pipelines. Shipped end-to-end systems serving real banking workflows — credit scoring engines, multi-template document parsing (29 formats, 90%+ accuracy, 95% manual workload reduction), and KYC automation APIs running at sub-2s latency. Comfortable owning problems from business framing through production deployment and post-launch optimization.

SKILLS

Languages: Python, Java, SQL, PL/SQL

ML & AI: Machine Learning, Deep Learning, NLP, Computer Vision, Time Series, LLM Fine-tuning (QLoRA), RAG, Prompt

Engineering, Model Compression (ONNX, Quantization), SHAP/Explainability

Frameworks & Libraries: PyTorch, TensorFlow, Hugging Face Transformers, scikit-learn, OpenCV, MMDetection, PaddleOCR, FastAPI

MLOps & Cloud: Azure ML, Docker, Kubernetes, Airflow, Git, CI/CD

Data Engineering: Apache Spark, Kafka, SQL, Data Warehousing, ElasticSearch, Kibana, Power BI

PROFESSIONAL EXPERIENCE

AI / ML Engineer · Talaty · Casablanca, Morocco | 07/2024 – Present

Statistical Credit Scoring Pipeline for Loan Solvency — Python · PySpark · Azure ML · Optuna · MLOps · SHAP

• Designed and deployed an end-to-end credit scoring pipeline (model AUC 0.81, Gini 62%) powering real-time loan eligibility decisions for banking clients.
• Engineered behavioral features from time-series bank account data (daily balances, cash flow patterns) at scale using PySpark.
• Applied feature selection to remove unstable and high-variance predictors, improving generalization and reducing overfitting.
• Built a dynamic hyperparameter optimization framework with Optuna; trained and tuned multiple ML models on Azure ML.
• Applied probability calibration (Platt scaling, isotonic regression) to align predicted default probabilities with realized default rates, enabling reliable use as inputs for risk-based pricing, expected loss estimation, and approval threshold optimization.
• Integrated SHAP-based explainability with GenAI to auto-generate human-readable justifications for each credit decision, supporting regulator and customer-facing transparency.
• Performed reject inference by scoring historical declined applicants with the trained model, quantifying both the volume of creditworthy candidates previously rejected and the incremental portfolio profit the bank would have realized had the model driven approval decisions from inception.

FinDoc Vision AI — Automated Bank Statement Parsing System — Python · Vision-Language Models · LLM Fine-tuning · Docker ·

FastAPI

• Designed and deployed a parallelized, low-resource AI pipeline for multi-template scanned bank statement extraction across 29 document formats, achieving 90%+ accuracy.
• Reduced manual processing workload by 95% and significantly shortened loan decision timelines.
• Fine-tuned transformer-based vision-language architectures (Table Transformer) to generalize across diverse document layouts.
• Trained a high-performance table detection model using MMDetection at 99% precision; applied ONNX quantization for low-latency inference in low-resource environments.
• Fine-tuned LLMs (Qwen, LLaMA) with QLoRA and engineered structured prompts (instruction design, few-shot) to correct OCR errors and improve textual consistency.
• Exposed the pipeline via FastAPI with Docker containerization and CPU-optimized inference; implemented parallelized processing and distributed inference for throughput.
• Owned full delivery lifecycle: business framing, data preparation, modeling, deployment, and performance optimization.

Automated Bank Check Parsing System — Python · YOLOX · PaddleOCR · TrOCR · Vision-Language Models · Docker

• Built a Dockerized AI-powered API for automated extraction of Moroccan bank check data, achieving 99% extraction accuracy at 2.5s average inference latency.
• Developed a synthetic data generation pipeline producing realistic Moroccan check images, improving model robustness and reducing reliance on manually labeled data.
• Fine-tuned YOLOX for high-precision detection of key check fields (RIB numbers, account holder names) under varied real-world conditions.
• Fine-tuned TrOCR for accurate recognition of RIB digits, significantly improving performance on structured financial identifiers.
• Quantized TrOCR and compressed the PaddleOCR pipeline to reduce inference latency, memory footprint, and deployment cost while preserving accuracy.
• Engineered a multiprocessing-based inference architecture for real-time, horizontally scalable workloads.

Moroccan ID Card Data Extraction API — Python · OpenMMLab · Synthetic Data · Docker · FastAPI

• Shipped a containerized AI API for automated data extraction from both sides of Moroccan ID cards at 95% accuracy and sub-2s response time, accelerating KYC in core banking systems.
• Collected, prepared, and labeled multi-source datasets for robust model training under diverse imaging conditions.
• Trained oriented object detection models to localize ID cards and key identity fields.
• Parallelized the text recognition pipeline using a compressed PaddleOCR recognizer for high-throughput real-time OCR.
• Led requirements discussions with stakeholders, translating business needs into API specifications and delivery roadmaps.

Data Science Intern · Talaty · Casablanca, Morocco | 02/2024 – 06/2024

Low-Resource Bank Statement Data Structuring Pipeline — Python · ML System Design · Machine Learning

• Contributed to a machine learning pipeline extracting and structuring data from diverse bank statement formats, achieving 86% accuracy at sub-1-second response time.
• Led requirement analysis, feasibility assessment, and system design to ensure robust, scalable implementation.
• Built a Python-based ML pipeline for data restructuring.
• Authored detailed technical documentation covering architecture, design patterns, algorithms, and data workflows.

Data & BI Engineer Intern · SINONE · Rabat, Morocco | 06/2023 – 09/2023

Data Engineering Platform for Strategic Decision-Making — Python · SQL · ElasticSearch · Kibana · Data Warehouse

• Designed a scalable data warehouse architecture for centralized analytics and KPI reporting.
• Developed automated ETL pipelines extracting, cleaning, transforming, and loading data from multiple operational databases into structured data marts.
• Built marketing-focused data marts surfacing customer behavior, campaign performance, and revenue trends.
• Deployed interactive BI dashboards enabling real-time visibility for management decision-making.

FREELANCE PROJECTS

Independent Engagement · Remote | 2025

Scalable Invoice Data Extraction & OCR Processing Platform — Python · FastAPI · Redis · PostgreSQL · Docker · NGINX · MinIO

• Architected an end-to-end asynchronous invoice processing platform using FastAPI, Redis queues, Docker worker fleets, PostgreSQL JSONB storage, and MinIO object storage for scalable GenAI-based data extraction workflows.
• Engineered a two-level parallel processing architecture supporting up to 16 concurrent client requests through multi- replica Docker workers and thread-based parallelism, enabling high-throughput real-time processing.
• Designed and deployed a production-grade REST API and NGINX gateway with secure large-file upload handling, request validation, queue orchestration, and end-to-end job lifecycle management.
• Built resilient job orchestration and recovery mechanisms: graceful worker shutdown, stuck-job detection, automatic re-queuing, retry tracking, and timeout-protected external API execution.
• Developed a Redis-backed rotating OpenAI API key management system with automatic rate-limit handling, dynamic cooldown management, and zero-downtime key rotation across distributed worker replicas.
• Implemented a multi-layer security architecture: IP allowlisting, API key authentication, HTTPS/TLS hardening, request rate limiting, fail2ban intrusion prevention, and secure runtime secret injection.
• Designed a CPU- and memory-aware autoscaling system that dynamically scales worker replicas based on Redis queue depth and host resource availability, improving workload stability and infrastructure efficiency.
• Deployed hybrid infrastructure combining containerized stateless services with bare-metal stateful services (PostgreSQL, Redis, MinIO) on Linux to optimize performance, reliability, and I/O throughput.
• Automated infrastructure security auditing using Trivy, Lynis, and custom validation scripts to identify vulnerabilities, validate service exposure, and strengthen production hardening.
• Authored comprehensive production documentation, deployment procedures, operational runbooks, and subsystem- level technical references to support long-term maintainability.

EDUCATION

State Engineer's Degree (Master's-equivalent), Data Science & Knowledge Engineering · École des Sciences de

l'Information (ESI) · Rabat, Morocco | 10/2021 – 06/2024

Preparatory Classes for Engineering Schools (CPGE) · Mathematics & Physics Track · Agadir, Morocco | 10/2019 – 06/2021

LANGUAGES

Arabic (native) · English (fluent)