Mohamed Ben Amor - AI & Data Science Student

Professional Experience

Data Analyst & NLP Intern

Paramedic 06/2025 – 08/2025

Paris

During my internship at Paramedic, I worked on two data-driven missions: building a trip attractiveness scoring system and developing an NLP pipeline for automatic incident detection in conversations. The internship combined data analysis (Excel) and NLP/machine learning (Python + Transformers) using real-world production data.

During my internship at Paramedic, a company specialized in coordinating medical transportation between hospitals, healthcare services, ambulance providers, and patients, I worked on two data-driven missions designed to improve operational decision-making and automate incident detection.

The internship combined data analysis (Excel) and NLP / machine learning (Python + Transformers) using real-world production data, with strong business constraints and confidentiality requirements.

Mission 1 - Trip Attractiveness Scoring (0–10)

Goal

The objective was to build an objective scoring system (from 0 to 10) to measure the attractiveness of transport trips, helping Paramedic compare and prioritize trips based on key operational factors such as price, duration, and constraints.

Approach

I followed a structured process:

Cleaned and prepared transport datasets in Excel, filtering out anomalies and extreme cases (e.g., unrealistic prices or durations).
Tested and compared three scoring strategies:
- Linear normalization
- Z-score based scoring
- Median-centered normalization (more robust to outliers)
Compared model outputs with professional ratings to validate relevance and stability.

Result

The median-based scoring method produced the best alignment with professional evaluations (lowest average gap and variability), and was delivered as a reusable Excel tool including automated formulas and visualizations.

Mission 2 - Incident Detection in Conversations (NLP)

Goal

A major operational challenge at Paramedic was identifying critical incidents in text conversations (delays, cancellations, missing drivers, patient issues). Manual detection was time-consuming and not scalable.

The goal was to build an end-to-end NLP pipeline to automatically classify conversations as:

Incident
Non-incident

Dataset & Preprocessing

Extracted real conversations from MongoDB using Python scripts.
Cleaned the dataset (removed automatic messages, duplicates, and low-signal texts).
Kept the dataset naturally imbalanced (more non-incidents than incidents) to reflect production conditions.
Built a structured training dataset in JSONL format, ready for model training and reuse.

Model Training

Fine-tuned CamemBERT (Hugging Face Transformers) for binary classification.
Used Google Colab GPU for training with stable hyperparameters (learning rate 2e-5, batch size 16, 2 epochs, weight decay).

Evaluation & Business Threshold Optimization

Initial results achieved:

~90% accuracy
Global F1 ≈ 0.91
Incident F1 ≈ 0.73

Because the business priority was to avoid missing incidents, I performed a threshold study (0.1 → 0.9) using precision/recall/F1 curves. A higher threshold (~0.90) significantly reduced false negatives while keeping precision acceptable.

Deliverables

Clean labeled dataset (JSONL)
Reusable Python scripts for extraction + dataset building
Trained CamemBERT model artifacts (weights/tokenizer/config)
Evaluation reports (confusion matrix, classification metrics, threshold analysis plots)

Tools & Methods Used

Excel Python (pandas, scikit-learn, matplotlib) MongoDB (pymongo) Hugging Face Transformers CamemBERT Google Colab (GPU)

Key Learnings

This internship strengthened my ability to:

Work with real operational data (cleaning, constraints, outliers)
Build complete ML/NLP workflows from scratch
Align technical metrics with business priorities (precision vs recall trade-off)
Deliver reusable, documented assets for future integration

Projects

These projects were developed as part of the "AI Clinic" module at Aivancity. In this course, students work in teams on real-world problems provided by partner companies. Each project follows a professional workflow, from understanding business needs to building a functional solution and delivering documented results.

Regulatory AI Assistant for Clinical Trials

AI / RAG Project

CERC (European Clinical Research Center)

Jan 2026 – Present

This ongoing project aims to build an internal "GPT-like" AI assistant to support clinical research teams in navigating complex regulations and internal procedures. The assistant uses a RAG approach to provide fast, reliable answers grounded in CERC's validated document corpus (SOPs and regulatory texts).

This ongoing project aims to build an internal "GPT-like" AI assistant to support clinical research teams in navigating complex and evolving regulations and internal procedures. The assistant must provide fast, reliable, and contextual answers while being strictly grounded in CERC's validated document corpus (SOPs and regulatory texts), without using any unverified external sources.

Context & Challenge

Clinical trials are governed by multiple frameworks such as ICH-GCP, EU MDR, GDPR, national regulations, and more than 300 internal SOP documents. Ensuring compliance during daily operational decisions is critical, but manually searching across these sources is time-consuming and error-prone.

Project Goals

Design an AI assistant architecture based on a closed, controlled knowledge base
Implement a robust Retrieval-Augmented Generation (RAG) approach for document-grounded Q&A
Ensure answers are traceable, explainable, and auditable (citations + source tracking)
Prevent hallucinations by enforcing strict document grounding
Provide a simple workflow to allow CERC to add/update/remove documents autonomously
Address governance, traceability, and compliance requirements

Planned Deliverables

A functional proof of concept (PoC) of the assistant
Technical documentation (architecture, indexing strategy, key design choices)
Tested real-world use cases with CERC teams
A proposed method for continuous updates of the reference corpus

Weather Trigger System

Data / Backend Project

Wellpack

Sep 2025 – Dec 2025

Built a decision-support system that allows business users to configure weather-based rules for multiple stores and automatically receive alerts when conditions are met. The system uses a Python Flask backend with Open-Meteo API integration, transforming business rules into automated backend logic.

This project was developed in collaboration with Wellpack and Aivancity. The goal was to build a decision-support system that allows business users (marketing or operations teams) to configure weather-based rules for multiple stores and automatically receive alerts when conditions are met.

The system is designed to be configurable, robust, and automated, transforming high-level business rules into concrete backend logic without requiring technical expertise from end users.

Project Objective

The project aims to:

Monitor weather conditions (temperature, rain, wind, UV, etc.) over a configurable time window (J-x → J+y)
Determine whether conditions are globally acceptable using tolerance logic
Automatically send summary emails when conditions are met, while avoiding notification spam

System Architecture

The solution is built around a Python Flask backend, structured into clear functional modules:

A Flask web application exposing HTTP routes for form submission and Excel import
A core business logic layer that validates inputs and orchestrates the analysis
A weather analysis module interacting with the Open-Meteo API
An email module handling SMTP notifications
A JSON-based cache used for persistence, state tracking, and anti-spam logic

This architecture allows the system to run both:

on demand (via the web interface), and
automatically (via a scheduler that reuses the last configuration).

Backend & Business Logic

A key part of the project was translating business-friendly inputs into reliable backend logic:

Parsing and validating user inputs coming from a web form or a structured Excel file
Converting textual business rules (thresholds, rain levels, time windows) into numerical conditions
Building pandas DataFrames to analyze weather data day by day
Applying a tolerance rule (up to one-third of days in a window may be non-compliant) to reduce false negatives

Weather Analysis

The backend interacts with the Open-Meteo API to:

Geocode store locations using postal codes
Retrieve forecast and historical weather data
Build detailed daily DataFrames containing weather values and compliance indicators
Decide whether a store is globally compliant across the configured time window

Automation & Anti-Spam Mechanism

To ensure reliability and usability:

A scheduler re-runs the analysis automatically using the last saved configuration
A failure counter system tracks consecutive non-compliant runs per store
After a defined threshold, notifications are temporarily silenced to avoid repetitive alerts
Email notifications are only sent when there is a meaningful state change

Deliverables

The project delivered:

A fully functional Flask backend
A web interface for configuration and result visualization
Automated email summaries
A reusable and documented backend architecture ready for extension

Key Learnings

This project strengthened my ability to:

Design structured backend architectures
Translate complex business constraints into algorithmic logic
Build robust data pipelines using APIs and pandas
Think in terms of reliability, automation, and real-world constraints rather than isolated features

Competitive Advertising & Recommendation System

AI / Competitive Intelligence Project

Egerine

Jan 2025 – Apr 2025

Developed a solution combining a questionnaire-based recommendation system for personalized advertising campaigns and a real-time competitive intelligence tool that automatically monitors competitor strategies using web scraping and NLP methods.

This project was developed in collaboration with Egerine, a company specialized in urban bicycle advertising. The objective was to strengthen Egerine's competitive positioning by combining personalized campaign recommendations with real-time competitor analysis.

The solution is structured into two complementary parts:

a questionnaire-based recommendation system for tailored advertising campaigns, and
a competitive intelligence tool that automatically monitors competitor strategies.

Part 1 - Questionnaire-Based Recommendation System (QCM)

Goal

The first component aims to help client companies quickly design a personalized advertising campaign by collecting key information through a structured interactive questionnaire.

How it works

The questionnaire collects data such as:

business type and campaign city
target audience and campaign objectives
budget, duration, and success metrics

Based on the answers, the system automatically generates:

a clear summary of the client's needs
tailored recommendations (message type, visuals, location strategy, operational tips)
a transparent calculation of daily exposure time derived from budget constraints

Business value

This approach improves:

personalization and client satisfaction
operational efficiency (faster campaign preparation)
transparency on budget usage
differentiation vs competitors through a structured and automated process

Part 2 - Real-Time Competitive Intelligence Tool

Goal

The second component aims to provide Egerine with a proactive competitive monitoring system by automatically collecting competitor data and generating actionable strategic recommendations.

Technical approach

The tool was designed to:

scrape competitor websites automatically using Python (requests, BeautifulSoup)
run periodic monitoring (planned every 30 minutes) to detect changes such as pricing, services, innovations, or promotions
extract key information using simple NLP methods (keyword extraction and regular expressions)
compare competitor features against Egerine's internal reference dataset (Excel file)

Outputs

The system produces:

structured comparisons of strengths/weaknesses per competitor
dynamic tables and visualizations exportable to Excel
operational recommendations (e.g., improving communication, adding interactive services, highlighting eco-impact)

Business value

This tool enables:

faster strategic reactions to market changes
continuous optimization of campaign offers
reduced manual effort and operational costs
clearer identification of differentiation opportunities

Key Learnings

This project improved my ability to:

Translate marketing/business needs into structured data workflows
Design recommendation logic based on user inputs
Automate competitor monitoring with web scraping and text processing
Deliver actionable insights through comparative analysis and reporting

AI & Data Science Student

Profile card

Professional Experience

Data Analyst & NLP Intern

Projects

Regulatory AI Assistant for Clinical Trials

Weather Trigger System

Competitive Advertising & Recommendation System

Education

Master's Degree in Artificial Intelligence & Data Science (Engineering Program)

French Baccalaureate (High School Diploma), Honors ("Assez Bien")

Skills & Languages

Technical Skills

Soft Skills

Languages

Get in Touch