Mohamed Amine Ben Charrada.

Medikair.tn distributes medical equipment and needed to move its dental range online for practices, clinics, and labs. The brief went past a standard storefront: the platform had to learn a practice's ordering habits, suggest restocks before they ran out, and propose equipment matched to the practitioner's speciality. I took it from specification through to delivery — a headless microservices architecture, a multi-level B2B catalogue with technical attributes and regulatory documents, a hierarchical ordering workflow for clinics, and an AI layer built in stages: content-based recommendations first, then collaborative filtering, K-Means behavioural segmentation, time-series reorder/churn prediction, and an NLP assistant grounded in the product documentation. It ships as two applications — the customer storefront at medikair.tn and the administration panel at admin.medikair.tn — both deployed locally on a self-hosted server rather than cloud infrastructure.

• —Hybrid recommender (content-based + collaborative) driving cross-sell in cart, up-sell on product pages, and predictive restock emails
• —K-Means customer segmentation on real purchase behaviour rather than declared status
• —Time-series prediction of consumable stock-out dates to trigger proactive reorder alerts
• —B2B ordering tunnel: persistent cart, CSV/Excel bulk order, one-click reorder, clinic-level validation workflow
• —Faceted search with typo tolerance and medical synonyms (Elasticsearch/Algolia class)
• —MLOps pipeline: ETL over historical orders and clicks, periodic retraining, models served over REST
• —Security & compliance: TLS 1.3, AES-256 at rest, GDPR, MFA for admin accounts, HDS-grade hosting where health data is involved

• — 2 applications — storefront + admin
• — 3-click consumable reorder flow
• — 26-week delivery, spec to launch

A full computer-vision system for automated match analysis. TrackNet — a U-Net-based architecture — detects and tracks the ball across sequential frames, holding up through fast motion, occlusion, and changing light. A custom CNN (CourtKeypointNet) locates the 14 court keypoints, which gives automatic calibration and perspective correction so ball positions map to real court coordinates. Faster R-CNN with a ResNet50-FPN backbone handles player detection, and a CatBoostRegressor predicts bounce locations from trajectory, velocity, and position. On top of that sits a statistics layer: shot classification (forehand, backhand, serve, volley), rally length, shot-speed estimation, court coverage, heatmaps, and positional density.

• —TrackNet (U-Net) ball detection and tracking with temporal context across frames
• —Redesigned CourtKeypointNet for 14-point court keypoint estimation, +8% accuracy
• —Faster R-CNN (ResNet50-FPN) player detection
• —CatBoostRegressor bounce detection from trajectory analysis
• —Real-time 1280×720 pipeline, CUDA-accelerated, Albumentations for dataset augmentation
• —Awarded Best Project of the Semester

• — 95%+ ball detection accuracy
• — +8% court keypoint accuracy
• — 14 court keypoints
• — 1280×720 real-time

My master's research thesis. The system decomposes diagnostic workflows across specialised agents orchestrated with CrewAI, rather than leaning on one monolithic LLM. Diagnosis runs on a monomodal NLP approach built on ClinicalBERT, fine-tuned for clinical text classification, reaching 98.66% test accuracy and addressing the performance stagnation of traditional LLMs on this task. A Reinforcement Learning with Human Feedback loop folds clinician corrections back into retraining, so the system improves in service. The treatment agent uses Retrieval-Augmented Generation to ground recommendations in verified medical literature, cutting hallucinations and improving factual reliability, and a Diagnostic Confidence Scoring System quantifies model reliability so a clinician can see how much weight an output deserves.

• —CrewAI multi-agent orchestration for diagnosis and treatment workflows
• —ClinicalBERT fine-tuned for clinical text classification — 98.66% test accuracy
• —RLHF loop integrating clinician feedback into retraining for a self-improving system
• —RAG grounding in verified medical literature to reduce hallucination
• —Diagnostic Confidence Scoring System for explainability and transparency
• —Interactive Streamlit application with patient history and statistical dashboards

• — 98.66% test accuracy
• — 10-month research thesis

Built end-to-end during a one-month internship at Anypli. Random Forest classifiers score transactions for anomaly against amount, type, and history, and the model is served to a cross-platform Flutter application through a Python REST API. The app lets agents enter, review, and analyse transactions live, with alerts firing the moment a transaction looks suspicious and a three-tier risk classification (low / medium / high). Firebase handles the real-time database and secure authentication; the dashboard covers transaction statistics, detection history, and system performance.

• —Random Forest anomaly detection on financial transaction data
• —Flutter mobile app with Provider state management and Material Design
• —Python REST API bridging the ML model and the mobile client
• —Firebase real-time database and secure authentication
• —Automated fraud alerts with three-tier risk scoring
• —UI/UX for fraud alerts and transaction visualisation

• — 1-month internship
• — 3-tier risk classification
• — Real-time scoring

A collaborative management platform giving a granular, analytical view of project health. The Spring Boot back end models the full organisational structure — projects, poles, personnel, tasks, dependency links — through JPA/Hibernate and exposes it over a RESTful API. A dedicated KPI service computes delay rate, deadline compliance, daily productivity, and schedule variance, with scheduled jobs updating overdue task statuses automatically. The Angular front end delivers interactive dashboards, a DHTMLX Gantt chart for detailed planning and dependency creation, and Chart.js analytics for project health monitoring. Optimistic locking on the Link entity keeps concurrent dependency edits safe.

• —Project, pole, personnel, and task lifecycle modelling via JPA/Hibernate
• —Automated KPI service: delay rate, deadline compliance, daily productivity, duration variance
• —DHTMLX Gantt chart with hierarchical tasks and dependency links
• —Chart.js analytics dashboards for project health and trend analysis
• —Scheduled jobs for automatic overdue-task status updates
• —Optimistic locking (@Version) for concurrent dependency updates

• — 2-month internship
• — 4 automated KPIs

A complete e-commerce system designed for rapid redeployment across clients. A central Node.js/Express RESTful API over MongoDB handles products, inventory, orders, reviews, and administrators, with schema modelling for complex relationships plus business logic for stock updates and order status. Authentication runs on a dual strategy — Firebase (Google) for customers, JWT and Bcrypt.js for admins — layered with three-level role-based access control (Admin, Provider, Client) and precise CRUD authorisation schemas. Stripe covers payments, Cloudinary hosts product imagery, and Formspree handles feedback. The React storefront ships product filtering and sorting, cart, checkout, order tracking, and reviews.

• —Decoupled MERN architecture: backend API, admin dashboard, storefront
• —Dual authentication — Firebase for customers, JWT + Bcrypt.js for admins
• —Three-level RBAC with per-resource CRUD authorisation schemas
• —Stripe payments with webhook handling, Cloudinary image hosting, Formspree feedback
• —Responsive React storefront: filtering, sorting, cart, checkout, order tracking, reviews
• —Admin dashboard for product CRUD, order status, and user administration

• — 3 applications
• — 4-month project
• — Multi-tenant ready

A full MERN e-commerce build for ArtBois, a wood and timber supplies retailer. The system ships as two deployed applications sharing one Node.js/Express API over MongoDB: a React storefront at artbois.tn where customers browse and order, and a separate React admin panel at admin.artbois.tn where the client manages the catalogue, stock, and incoming orders. Both applications are deployed and maintained on a self-managed Hetzner server.

• —Two-application split — customer storefront and admin back-office on separate subdomains
• —Shared Node.js/Express REST API backed by MongoDB
• —React front ends for both the shop and the management dashboard
• —Self-hosted deployment, domain, and server setup on Hetzner

• — 2 applications
• — Live in production
• — Self-hosted

A full MERN e-commerce build for Darista, a home products retailer. As with ArtBois, the system is split into two deployed applications over a shared Node.js/Express API and MongoDB: a React storefront at darista.tn for browsing and ordering, and a React admin panel at admin.darista.tn for catalogue, stock, and order management. Deployed and maintained on a self-managed Hetzner server.

• —Two-application split — customer storefront and admin back-office on separate subdomains
• —Shared Node.js/Express REST API backed by MongoDB
• —React front ends for both the shop and the management dashboard
• —Self-hosted deployment, domain, and server setup on Hetzner

• — 2 applications
• — Live in production
• — Self-hosted