Open to internships & research collaborations

Anas Meftah

Applied ML Engineer · AI Engineer · Trustworthy AI Builder · Neural Network Verifier

Making AI reliable, one bound at a time.

I'm a Data Science engineering student at the Faculty of Science of Tunis building machine learning systems that hold up outside the notebook — from drift monitoring and RAG pipelines to certified robustness of neural networks.

View projects Download CV

Tunis, Tunisia

python

import numpy as np


def linear_bounds(W, b, lower, upper):
    """Sound bounds for y = Wx + b over a box."""
    W_pos, W_neg = np.maximum(W, 0), np.minimum(W, 0)
    new_lower = W_pos @ lower + W_neg @ upper + b
    new_upper = W_pos @ upper + W_neg @ lower + b
    return new_lower, new_upper


def relu_bounds(lower, upper):
    return np.maximum(lower, 0), np.maximum(upper, 0)

# ∀ x in the ε-ball: lower ≤ f(x) ≤ upper

mainanesmeftah/IBP_verifier

Python
PyTorch
Scikit-learn
NumPy
Pandas
SciPy
α,β-CROWN
ONNX
MLflow
Docker
LangChain
MCP
Streamlit
SQL
Git
Linux
Jupyter
LaTeX

About

From the math up.

I'm pursuing a National Engineering Diploma in Data Science at the Faculty of Science of Tunis (University of Tunis El Manar), graduating in 2028 — after an integrated engineering preparatory cycle with honors and a Mathematics Baccalaureate with highest honors. I tend to approach machine learning from the math up.

As an applied ML engineer, I care about what happens after the notebook: data pipelines, evaluation, monitoring, and tooling. I've built a feature-drift monitoring pipeline with Kolmogorov–Smirnov tests implemented from scratch in NumPy, an MCP server that grounds LLM answers in PDF documents through a retrieval pipeline, and a from-scratch Vision Transformer in PyTorch.

My focus is trustworthy AI. In July 2026 I joined the Digital Research Center of Sfax (CRNS) as a research intern in neural network verification, evaluating the certified robustness of CNNs on the VNN-COMP CIFAR2020 benchmark with α,β-CROWN. Alongside it, I wrote my own Interval Bound Propagation verifier to understand these algorithms from first principles.

~/now

$ cat now.md

Studying: Engineering Diploma in Data Science @ FST · 2025 – 2028
Recently: Research intern, neural network verification @ CRNS · July 2026
Building: IBP Verifier — symbolic interval propagation for tighter bounds

$ locale --list

Arabic: Native
English: Professional
French: Professional

ML Engineering
End-to-end pipelines — preprocessing, training, evaluation, and the monitoring that keeps models honest in production.
Trustworthy AI
Robustness and formal verification of neural networks: bounding what a model can do, not just measuring what it did.
LLM Applications
Retrieval pipelines and MCP integrations that ground LLM answers in real documents.
Statistical Learning
Hypothesis testing, stochastic modeling, and time series — the statistics underneath reliable ML.

Experience

Research meets engineering

From verifying neural networks at a national research center to applied ML for a cybersecurity R&D team.

Education & more

ResearchJul 2026
Research Intern — Neural Network Verification
The Digital Research Center of Sfax (CRNS) · Sfax, Tunisia
- Evaluated the certified robustness of convolutional neural networks on the VNN-COMP CIFAR2020 benchmark with α,β-CROWN — 147 properties under L∞ perturbations (ε = 2/255 and 8/255).
- Reproduced the official competition verdicts with 93.2% agreement (137 / 147), including all 103 certified-safe (UNSAT) instances; divergences clustered near the 300 s timeout.
- Automated parsing and analysis of α,β-CROWN verification logs in Python, extracting verdicts, runtimes, and Branch-and-Bound statistics into structured reports.
- Investigated Branch-and-Bound behaviour and robustness under different perturbation budgets, and built a PySide6 GUI to drive verification experiments from one place.
- α,β-CROWN
- Certified Robustness
- VNN-COMP
- Python
- PySide6
R&D2024 — Present
R&D Member — Cybersecurity & Data
Securinets FST Student Club · Tunis, Tunisia
- Processed and cleaned the NSL-KDD intrusion detection dataset (~125,000 records, 42 features), performing EDA and correlation analysis to guide feature selection.
- Trained a Random Forest classifier with Bayesian hyperparameter tuning (skopt), achieving competitive F1 scores across attack categories; evaluated via per-class precision, recall, and confusion matrices.
- Delivered a 3-page interactive Streamlit dashboard (Overview · Feature Importance · Model Comparison) enabling non-technical stakeholders to explore model results.
- Machine Learning
- Cybersecurity
- Streamlit
- Scikit-learn

Projects

Selected work

Systems built to be measured, monitored, and verified — from certified robustness to drift monitoring and LLM tooling.

All projects

Trustworthy AI2026
IBP Verifier
A from-scratch neural network verifier that computes sound output bounds for ONNX models under L∞ perturbations, using interval and symbolic bound propagation.
- Python
- NumPy
- ONNX
- pytest
SourceCase study
Trustworthy AI2026
α,β-CROWN Reproducibility Study
An independent reproduction of α,β-CROWN on the VNN-COMP CIFAR2020 benchmark: 147 robustness properties verified locally, with 93.2% agreement with the official verdicts.
- Python
- α,β-CROWN
- Pandas
- Matplotlib
- Jupyter
SourceCase study
MLOps2024 – 2025
Feature Drift Detection
A statistical monitoring pipeline that flags feature drift between reference and production data, with Kolmogorov–Smirnov tests implemented from scratch in NumPy.
- Python
- NumPy
- Pandas
- Scikit-learn
- Streamlit
SourceCase study
LLMs & RAG2025
MCP Server for PDF Q&A
A Model Context Protocol server that connects PDF documents to LLMs through an automated retrieval pipeline — extraction, chunking, embedding, and similarity search.
- Python
- MCP
- LangChain
- Embeddings
- Vector Search
SourceCase study

Skills

Technical toolkit

The languages, frameworks, and methods I reach for — from the numerical stack to verification tooling.

Languages
Python first, with the SQL and shell to move data around.
- Python (Advanced)
- SQL
- Bash / Shell
- Java
Scientific Computing & ML
The numerical stack I build models and experiments on.
- NumPy
- Pandas
- SciPy
- PyTorch
- Scikit-learn
- Matplotlib / Seaborn
Trustworthy AI
Methods and tools for bounding and certifying model behaviour.
- α,β-CROWN
- Interval Bound Propagation
- Symbolic Bound Propagation
- Branch-and-Bound
- ONNX
- VNN-LIB
LLMs & Applied AI
Grounding language models in real documents.
- RAG Pipelines
- MCP
- LangChain
- Embeddings & Retrieval
- Streamlit
MLOps & Statistics
Keeping models measurable once they ship.
- Drift Detection (KS)
- Hypothesis Testing
- MLflow
- Model Evaluation
- Bayesian Tuning (skopt)
Tools & Workflow
The everyday kit for reproducible work.
- Git
- Linux CLI
- Docker
- Jupyter
- LaTeX
- PySide6

Research

Research interests

Where my curiosity points — anchored by hands-on work in neural network verification.

Explore research

01
Neural Network Verification
Formally bounding what a network can output for every input in a region: interval and symbolic bound propagation, linear relaxations, and Branch-and-Bound search.
02
Trustworthy AI
ML systems whose behaviour holds up under distribution shift and adversarial perturbation — backed by monitoring and guarantees rather than hope.
03
Machine Learning
Learning algorithms end to end, from classical models to transformers — rebuilt from scratch whenever that is the best way to understand them.
04
Statistical Learning
Hypothesis testing, inference, and the statistical foundations that turn model outputs into defensible conclusions.
05
Scientific Computing
Vectorized numerical methods in NumPy and SciPy — the machinery behind fast bound propagation, signal processing, and statistical tests.
06
Complex Systems
How global behaviour emerges from many interacting parts — networks, stochastic dynamics, and the graph-theoretic tools used to study them.

Open Source

Let's build AI you can trust.

Open to internships, research collaborations, and interesting ML engineering problems. Based in Tunis, Tunisia.

Get in touch Email me

Anas Meftah

From the math up.

ML Engineering

Trustworthy AI

LLM Applications

Statistical Learning

Research meets engineering

Research Intern — Neural Network Verification

R&D Member — Cybersecurity & Data

Selected work

IBP Verifier

α,β-CROWN Reproducibility Study

Feature Drift Detection

MCP Server for PDF Q&A

Technical toolkit

Languages

Scientific Computing & ML

Trustworthy AI

LLMs & Applied AI

MLOps & Statistics

Tools & Workflow

Research interests

Neural Network Verification

Trustworthy AI

Machine Learning

Statistical Learning

Scientific Computing

Complex Systems

More on GitHub

GUI-abcrown

abcrown-result-parser

deep-learning-roadmap

Kaggle

AutoEncoders

Ecommerce-Analysis

Let's build AI you can trust.