Machine learning and deep learning models
Prediction, classification, forecasting and anomaly detection on your data, for retail, industry, health or finance.
Projects: DataBot, ThreatLens, EEG
Open to full-time roles and freelance projects, remote or in Tunis
Data scientist and AI engineer · ENIT graduate, 2026
I build AI assistants, predictive models and computer-vision tools that work on real data.
I measure honestly what my models are worth before showing them.
About
Modelling and data science engineer, graduated from the National Engineering School of Tunis in 2026. I like complete projects: start from raw data, build a model or an agent, evaluate it on data it has never seen, then ship it in a tool someone actually uses.
My projects span classical machine learning, deep learning for vision, signal processing and generative AI: agents, RAG and LLM-as-judge evaluation.
Services
Each service is backed by projects you can open, with their measured results and limits.
Prediction, classification, forecasting and anomaly detection on your data, for retail, industry, health or finance.
Projects: DataBot, ThreatLens, EEG
AI assistants that answer from your documents or data (RAG) and agents that automate multi-step tasks.
Projects: Explain My Model, DataBot
Detection and classification on images and video, from data cleaning to real-time inference.
Projects: Facial emotion recognition
Projects
Each project shows its measured results, its method and its limits.
2 video demosAsk questions about your business data in plain language and get the answer with its chart, or run a full data science project together, approving each step, with everything delivered as a notebook and a PDF report.
An AI assistant that connects to databases and files, writes and runs the code, builds and evaluates models. It follows CRISP-DM and only moves to the next phase once the user approves. Ready for public deployment: open accounts, each user's own AI key, isolated code.
Video demoRecognises the emotion on each face in a video (happiness, surprise, anger…) and shows how sure it is, faster than the video plays.
A pre-trained ResNet-18 recognises 7 emotions on every face in a video, in real time. I first removed 7,976 doubtful images, including 1,916 copies shared between training and test, then measured on a clean test set: 82.7% accuracy.
Results per patientAutomatically spots epileptic seizures in brain-signal (EEG) recordings.
One model per patient on the CHB-MIT database: filtering, signal features, channel selection with Boruta and a threshold tuned not to miss seizures. 31 of 32 seizures detected.
Reads the SAS, Python and R code of a bank's credit risk models, even code built on SAS Risk Dimensions, a proprietary engine with no public documentation, and writes their technical and regulatory documentation, checked automatically and then approved by an expert.
Final-year project at PwC: eight AI agents read the SAS, Python and R code of credit risk models and write their IFRS 9 and Basel III documentation. Automatic checks, an LLM judge and expert review: 0.96 faithfulness to code.
Watches a website's traffic and flags suspicious visitors (vulnerability scans, password attacks, bots), showing where they come from and sending an alert by email.
Unsupervised detection of scans, attacks and bots in 1.83 GB of web logs. Isolation Forest scores each address every 5 minutes; the ThreatLens app shows a map and sends PDF alerts. F1 of 0.92 on a simulated test set.
Approach
I start by setting what matters and how to measure it: missing an epileptic seizure costs more than a false alarm, so recall comes first; in my final-year project, five measurable success criteria were set before the first line of code.
Before modelling, I look for what would distort the result: 1,916 images shared by training and test removed, customers counted twice, incomplete months, amounts inflated by joins.
An isolated test set used only once, real time order when it matters, comparison with reference models, confidence intervals. Settings are chosen on validation, never on the test set.
I go all the way to the tool people use: a dashboard with alerts, a platform for experts, PDF reports, a notebook. Every deliverable says where it can be trusted and where it cannot.
Experience
February to July 2026 · Tunis
Final-year engineering project in Risk Services Maghreb: a multi-agent generative AI platform that reads the code of credit risk models and writes their IFRS 9 and Basel III documentation, checked automatically then approved by an expert. View project.
July to August 2025 · remote
Unsupervised attack detection in 1.83 GB of web logs, and ThreatLens, a monitoring application with a map and PDF email alerts. View project.
AI Engineer for Data Scientists Associate, DataCamp, 11 December 2025. View the certificate
Skills
Used in: Emotions
Used in: DataBot, Explain My Model
Used in: DataBot, Explain My Model, Anomalies
Used in: Explain My Model, Anomalies, EEG, DataBot
Contact
I am open to full-time roles and freelance projects, remote or in Tunis. The easiest way is to email me: I reply quickly.
emnatounssi@gmail.com