Available for a 4–6 month end-of-studies internship · Feb 2027

Chehin Chaaben.

> Final-year ICT engineering student at SUP'COM. I ship production-grade AI models, the data pipelines that feed them, and the backends that put them in front of users.

View selected work Get in touch GitHub
  • Python
  • PyTorch
  • scikit-learn
  • pandas
  • SQL
  • FastAPI
  • Docker
  • LLMs
/about

Grounded, measured, shipped.

I'm a final-year ICT engineering student at SUP'COM Tunis. I got there through two years of preparatory classes, ranking 17th nationally in the engineering entrance exam.

I work where AI meets software engineering, from classic machine learning to deep learning and language models. What I enjoy most is taking a messy real-world problem and turning it into something people can actually use.

Internships, school projects, hackathons and a few entrepreneurial projects taught me to ship under real constraints. Next, I want to build AI products that take research out of the notebook and into production.

Solving real needsUsing technology where it makes a real difference, not just because it's new.
Understanding the businessListening to what a team actually needs, then building what serves it.
Automation & useful toolsRemoving repetitive work so people can focus on what matters.
Reliable dataMaking data trustworthy first, then getting much more value out of it.
17thnational rank, engineering entrance exam
3hackathon podiums
9projects built, solo and in teams
3industry experiences: EY, Sofrecom, ATB
/experience

Where I've been plugging the wires in.

  1. 2026 sep → now

    EY Tunisia · Tutored project, Data & AI

    Tunis, Tunisia

    Financial document extraction platform for audit: PDF, scans, Word, Excel; French, English, Arabic.

    • Designed and developed an end-to-end financial document extraction platform for audit, running 13 AI models locally and reading over 80% of tables correctly.
    • Benchmarked more than 20 extraction methods on more than 2,000 pages of financial documents.
    • Combined OCR and vision-language reading verified against document totals, reaching over 98% text fidelity and 96% of amounts on the correct line on degraded pages.
    • Built a quality check that flags suspicious tables to the auditor, catching over 83% of simulated errors with under 1% false alarms.
    • Python
    • PyTorch
    • Hugging Face
    • Transformers
    • ONNX Runtime
    • PaddleOCR-VL
    • llama.cpp
    • OCR
    • pytest
  2. 2026 jun → aug

    Sofrecom Tunisia · Data & AI intern

    Sfax, Tunisia

    Partner-onboarding support platform.

    • Automated partner onboarding tracking, replacing 5 manual lookups across email, Jira tickets and tracking pages with a single view and zero manual entry.
    • Designed 5 specialized agents on local SLMs served with Ollama, each with its own role and context, plus a chatbot that queries the entire partner portfolio.
    • Built an anti-hallucination harness covering both the agents and the chatbot, keeping every generated answer grounded in the partner data.
    • Engineered a traceable, reversible event-driven architecture with event sourcing, so every status change can be audited and undone.
    • Integrated a remote inference engine under confidentiality constraints and set up monitoring of the data flow and agent performance.
    • Python
    • LLM/SLM
    • Prompt engineering
    • SQLite
    • ETL
    • Ollama
    • REST
    • Event sourcing
    • Jira
  3. 2025 jul → aug

    Arab Tunisian Bank (ATB) · Summer intern, IT Security

    Tunis, Tunisia

    Two-month summer internship in the IT security team of a Tunisian retail bank.

    • Reviewed the controls, audits and monitoring that keep a retail bank's core banking infrastructure secure and compliant, during a two-month internship in the IT security team.
    • Studied the bank's cybersecurity procedures, data management practices and service continuity processes, where traceability and least privilege come first.
    • IT security
    • Banking
    • Monitoring
    • Compliance
/education

Where the foundations come from.

  1. 2024 → 2027

    SUP'COM · University of Carthage · Engineering degree, ICT

    Tunis, Tunisia

    Final-year student. 4–6 month end-of-studies internship starting February 2027.

  2. 2022 → 2024

    IPEIS · Preparatory cycle, Technology

    Sfax, Tunisia

    Classes préparatoires. Ranked 17th nationally.

/selected_work

Things I've built, solo and in teams.

A short tour of what I've been shipping: medical imaging, multimodal search, market and demand forecasting, energy markets, health wearables, and assistants that actually help.

  1. 01 feb → may 2026

    Lung nodule detection on 3D chest CT

    An end-to-end pipeline for automatic detection of pulmonary nodules on volumetric scans, from lung segmentation to candidate scoring.

    • Trained nnU-Net-based segmentation models on 3D chest scans in PyTorch, reaching a segmentation Dice score of 0.72 ± 0.08.
    • Developed an end-to-end detection pipeline, from lung segmentation and 3D reconstruction to candidate scoring, that flags suspicious regions automatically.
    • Evaluated the detection stage at an F1 score of 0.58, scoring nodule candidates with 3D CNN classification to support malignancy assessment.
    • PyTorch
    • nnU-Net
    • 3D CNN
    • CT imaging
    • Python
    View code ↗
  2. 02 feb 2026

    KANZ: Tunisian stock market intelligence

    1st place · IHEC CodeLab 2.0Team project

    Makes the Tunis Stock Exchange (BVMT) easier to read: forecasts, news sentiment and anomaly signals in one assistant.

    • Predicted the direction of BVMT price moves with an XGBoost model, reaching about 70% directional accuracy on market data.
    • Engineered time-series features from BVMT market data and external sources to train LSTM and XGBoost forecasters.
    • Automated OCR extraction of BVMT market documents, turning them into structured inputs for the forecasting pipeline.
    • Combined forecasts, news sentiment and anomaly signals into one decision-support assistant that won 1st place at IHEC CodeLab 2.0.
    • Python
    • XGBoost
    • LSTM
    • OCR
    • Time series
    View code ↗
  3. 03 mar 2026

    BMO: a cognitive assistant for your file system

    Team project

    A document-research assistant that watches your file tree and keeps a vector index fresh in the background. Ask in plain language, get the file in seconds.

    • Implemented a RAG pipeline that chunks, embeds and indexes the local file system into a Qdrant vector store for retrieval-augmented answers.
    • Designed event-driven incremental ingestion with Watchdog that updates the vector index in real time as files change, with no manual re-indexing.
    • Delivered a conversational search interface that finds a file from a natural-language description in about 3 seconds.
    • RAG
    • Qdrant
    • Watchdog
    • LLM
    • Python
    View code ↗
  4. 04 sep → oct 2026

    Green Wallet: decentralized peer-to-peer energy trading

    Team project

    Neighbors trade energy directly, with no central utility in between: every 30 minutes, a double auction matches surplus with need at one fair price, with payments and green certificates on a Hedera-style ledger.

    My part: The market logic and everything data: datasets, ETL pipelines and the test scenario.

    • Designed the double-auction market logic that clears peer-to-peer energy trades every 30 minutes, serving own needs first, then batteries, the auction and the grid, so no household is ever cut off.
    • Built 3 ETL pipelines in Python and pandas that rebuild real household needs from public datasets for Germany, Australia and Ireland, merged into one format and timeline.
    • Wrote the end-to-end test scenario for the trading platform and validated it, with all 5 of 5 checks passing.
    • Python
    • pandas
    • ETL
    • TypeScript
    • Blockchain
    • Simulation
    View code ↗
  5. 05 jan 2026

    BikeShare demand forecasting

    Team project

    Forecasts bike-share demand in Washington D.C. and explains it in a dashboard, then predicts real 2026 dates with live weather.

    My part: The machine-learning side: models, training, evaluation and the 2026 forecasting adaptation.

    • Trained an XGBoost demand model for Washington D.C. bike-share, raising R² from 0.626 for a seasonal baseline to 0.927, at 12.7% MAPE.
    • Engineered 30+ time-series and weather features in scikit-learn and explained the model's main demand drivers with SHAP.
    • Extended the model to forecast 2026 demand from live weather data, feeding the team's forecasting dashboard.
    • Python
    • XGBoost
    • scikit-learn
    • SHAP
    • Time series
    • Feature engineering
    View code ↗
  6. 06 nov 2025

    Wellvia: AI health & nutrition companion

    4th place · IEEE CSTAM 2.0Team project

    An ESP32 wearable streams your vitals to a Flutter app; snap your meal and AI logs the nutrition.

    My part: All the development and the wearable schematic; the team helped with the submission.

    • Designed the ESP32 wearable schematic measuring heart rate, SpO2 and temperature, and streamed the vitals over Bluetooth Low Energy to the app.
    • Developed the Flutter app and its FastAPI and Firebase backend, raising three-level health alerts from live vitals.
    • Integrated a vision model that logs nutrition from meal photos and a local LLM assistant served with Ollama.
    • Delivered the full working prototype, from wearable to app, for a team of 4 that placed 4th at IEEE CSTAM 2.0.
    • Flutter
    • FastAPI
    • Firebase
    • ESP32
    • Ollama
    View code ↗
  7. 07 jan 2026

    Multimodal AI discovery platform

    Team project

    A product search engine where you type a sentence or drop an image, and results come back in near-real time, with recommendations tuned to your history.

    • Built a multimodal product search engine that matches text or image queries against a catalog with 512-dimension CLIP embeddings in Qdrant.
    • Indexed the embeddings with HNSW vector search and served it through FastAPI, giving near-instant cross-modal ranking.
    • Implemented personalized recommendations from search history and interaction patterns, plus gift ideas from a free-text description of a person.
    • CLIP
    • Qdrant
    • FastAPI
    • Python
    View code ↗

// more projects

ENSTAB

3rd place · Hack4UCAR 1.0Team project

University data turned into role-based dashboards, multilingual document search and plain-language SQL.

My part: The architecture, including user identification and access, and the AI features.

  • Designed the platform architecture, including user identification and role-based access to university data and dashboards.
  • Developed multilingual document search over university documents with pgvector embeddings in a FastAPI backend.
  • Implemented natural-language-to-SQL so users can query university data in plain language instead of writing SQL.
  • FastAPI
  • pgvector
  • Next.js
View code ↗

JINEN

Nursery management platform with AI child-development tracking, an LLM chatbot and vector search.

  • Built hybrid GraphRAG retrieval that combines a knowledge graph with vector search to give the LLM context for its reasoning.
  • Unified structured data and textual observations from multiple sources in MongoDB, covering 9 behavioral dimensions.
  • Created an LLM assistant that produces personalized recommendations and early alerts when concerning behaviors are detected.
  • Spring Boot
  • Angular
  • MongoDB
  • Knowledge Graph
  • GraphRAG
  • LLM
View code ↗
/focus_areas

What I can bring to your team.

Six areas I've worked in, each one backed by a project or an internship on this page.

Machine learning & deep learning

Designing, training and honestly evaluating models on real data, from images and 3D scans to tabular data.

  • PyTorch
  • scikit-learn
  • CNNs
  • Model evaluation

seen in: Lung nodule · Multimodal search

Document AI & retrieval

Turning PDFs, scans and spreadsheets into clean, searchable data, then answering questions grounded in it.

  • OCR
  • Vision-language models
  • Qdrant
  • GraphRAG
  • Hybrid search

seen in: EY · BMO

LLM applications & agents

Assistants and multi-agent systems built on language models, with clear roles, guard-rails and answers you can trace.

  • Ollama
  • LangGraph
  • MCP
  • Prompt engineering

seen in: Sofrecom · BMO

Backends & data pipelines

APIs, databases and ETL that turn a notebook into something a team can use, plus the monitoring to debug it.

  • FastAPI
  • SQLite
  • ETL
  • Event sourcing

seen in: Sofrecom · Green Wallet

Forecasting & anomaly detection

Time-series models that anticipate trends and flag what looks wrong, from stock prices to sensor data.

  • XGBoost
  • LSTM
  • Time series
  • Anomaly detection

seen in: KANZ · BikeShare

Full-stack & mobile apps

From API to screen: FastAPI and Spring Boot backends with Angular and Flutter front ends, wired to Firebase or MongoDB.

  • FastAPI
  • Spring Boot
  • Angular
  • Flutter

seen in: Wellvia · JINEN

/stack.yaml

Tools I reach for on autopilot.

Grouped the way I actually use them, not the way they look on a job post.

# ml & deep_learning

  • PyTorch
  • scikit-learn
  • XGBoost
  • TensorFlow
  • Transformers
  • Hugging Face
  • CNNs
  • LSTM
  • CLIP
  • nnU-Net
  • SHAP
  • ONNX Runtime

# data & analytics

  • pandas
  • NumPy
  • Polars
  • Jupyter
  • Matplotlib
  • Seaborn
  • Power BI
  • Power Query
  • ETL pipelines

# documents & rag

  • Knowledge graphs
  • GraphRAG
  • Qdrant
  • Embeddings
  • PaddleOCR-VL
  • Docling
  • EasyOCR
  • Tesseract
  • BM25

# llms & agents

  • LangChain
  • LangGraph
  • Multi-agent
  • MCP
  • Prompt engineering
  • Ollama
  • llama.cpp

# backend & apis

  • Python
  • SQL
  • FastAPI
  • Spring Boot
  • REST APIs
  • SQLite
  • MongoDB
  • Pydantic
  • Event sourcing

# apps & frontend

  • Flutter
  • Angular
  • Next.js
  • TypeScript
  • Firebase

# dev & infra

  • Git
  • GitHub
  • Linux
  • Docker
  • Bash
  • CI/CD
  • pytest

# also_familiar_with

  • Cloud principles
  • Big data
  • Security
  • Networks
  • C++
  • Embedded systems
/contact.sh

Let's build something
that measures up.

I'm actively looking for a 4–6 month end-of-studies internship starting February 2027 in AI engineering. Always happy to talk about machine learning, AI products, or any problem where data can help.

chahin.chaaben@supcom.tn LinkedIn GitHub