Final-year projects at Hydatis. Real products, real users.
16 end-of-studies internship subjects across vision-language models, video intelligence, IoT and autonomous agents, generative AI, fintech platforms, cybersecurity and data engineering. Each subject is attached to a Hydatis product or an ongoing R&D track, with a dedicated technical supervisor in Tunis.
Four steps, no email needed.
Applications are received directly in our recruitment system. You get an answer after review and, if shortlisted, a technical interview with the subject supervisor.
Read the 16 subjects below and choose the one that matches your specialty and your ambition.
Fill in the form at the bottom of this page with your CV (PDF) and why this subject motivates you.
A 45-minute discussion with the supervisor about the subject, your projects and your stack.
Internship agreement with your school, 4 to 6 months, Tunis office (hybrid possible).
16 subjects, 7 topics.
Every subject below is a real need of a Hydatis platform (Govora, CyberAware, Garantia, Solventia, LumniaSync) or of an active R&D collaboration.
1Multimodal AI & Vision-Language Models
Text + image (+ audio) sentiment model on social-media content about energy, pollution and carbon footprint, with explanations for decision makers. Alongside an industrial PhD (MOBIDOC) hosted by Hydatis.
Python, PyTorch, Transformers, PEFT/LoRA, Qwen2-VL / LLaVA, SHAP, Label Studio, WhisperApply for VLM-01
Reduce the energy cost and carbon footprint of VLM inference without losing accuracy: quantization, distillation, token pruning and an adaptive inference policy, with energy and CO2 measurement.
PyTorch, bitsandbytes, vLLM / llama.cpp, ONNX Runtime, CodeCarbon, Grafana, NVIDIA JetsonApply for VLM-02
Ingestion pipeline, sentiment time series per topic and region, correlation with open energy and emission data, interactive decision dashboard with explanation drill-down.
Python, Airflow, PostgreSQL + pgvector, statsmodels / Prophet, React or MetabaseApply for VLM-03
2Video Intelligence
Extend our construction-site prototype (IP cameras + YOLO) with a vision-language model that understands the scene, checks PPE compliance, estimates progress and writes the daily PDF report.
OpenCV, Ultralytics YOLO, Qwen2-VL / Florence-2, NestJS, React, PostgreSQL, MinIO, DockerApply for VLM-04
Detect business events on multi-camera streams (queues, crowding, empty shelves) and let managers query video archives in plain language, with incident reports and KPIs.
FastAPI, YOLOv8 / RT-DETR, CLIP / SigLIP, Qwen2-VL, Milvus or pgvector, Redis Streams, GrafanaApply for VLM-05
3Internet of Things & Autonomous Agents
Continue our summer 2026 work on autonomous agents (HVAC, lighting, occupancy, tariff) negotiating energy decisions, connect them to real sensors and actuators and measure savings on a pilot floor.
ESP32 / Raspberry Pi, MQTT, ThingsBoard, SPADE / LangGraph / CrewAI, Gymnasium, EnergyPlus, NestJS, ReactApply for IoT-01
Based on our LumniaSync solution for municipal lighting, design ML / LLM agents that supervise thousands of light points, detect failures, plan maintenance and adapt dimming to traffic and weather.
ThingsBoard, MQTT / LoRaWAN, scikit-learn, Prophet, LangGraph, Spring Boot, Angular, TimescaleDB, GrafanaApply for IoT-02
4Generative AI & AI Agents
Design a Model Context Protocol server over the Govora board-management platform so that AI assistants and n8n workflows can act on meetings, resolutions and documents with Keycloak permissions and a full audit trail.
TypeScript, NestJS, MCP SDK, Keycloak OAuth 2.1, PostgreSQL, Redis, Docker, Claude Desktop / n8n clientsApply for GenAI-01
RAG assistant over board documents and transcripts that drafts minutes in FR / EN / AR, extracts decisions and actions, tracks follow-up and answers with citations, with an on-premise LLM option.
NestJS, React, Whisper, Ollama / vLLM, pgvector or Qdrant, LangChain / LlamaIndex, RAGASApply for GenAI-02
Parse tender documents, build the compliance matrix, retrieve references and CVs from the company knowledge base and draft proposal sections under human review.
Python, LangGraph, Docling, FastAPI, pgvector, Odoo API, Claude API or open models, DOCX exportApply for GenAI-03
5FinTech Platforms: Garantia & Solventia
Scoring model with explainability, scorecard UI for bank analysts, fairness and drift monitoring, exposed as a scoring API integrated with Solventia.
Python, LightGBM / XGBoost, SHAP, FastAPI, PostgreSQL, React, EvidentlyApply for Fin-01
OCR and vision-language models on IDs, financial statements and invoices, consistency checks and a fraud-risk score integrated into the Garantia workflow.
Python, Docling, Tesseract, Qwen2-VL / Donut, anomaly detection, NestJS, PostgreSQL, MinIOApply for Fin-02
Monte-Carlo simulation of claims and leverage scenarios with a projection dashboard for fund managers, connected to the Garantia APIs.
Python, pandas, NumPy, Streamlit or React, PostgreSQLApply for Fin-03
6SecureTech & CyberAware
Campaign engine that adapts difficulty and themes to each employee’s results, LLM-generated content with guardrails, SCORM / xAPI learning paths and a maturity score per team.
NestJS, React, LLM content generation, SCORM 1.2 / xAPI, PostgreSQL, sandboxed email infrastructureApply for Sec-01
Log collection from our platforms, detection rules, alerting and a SOC dashboard on a lightweight open-source stack.
Wazuh or ELK, Falco, Kubernetes, Docker, Sigma rules, Grafana, n8n / Slack alertingApply for Sec-02
7Data Engineering
Profiling, validation rules, drift detection and lineage on a chain of ETL pipelines like those we run for our BI clients.
dbt, Great Expectations or Soda, Apache Airflow, OpenLineage / Marquez, PostgreSQL, Grafana, PythonApply for Data-01