# Kacem Mathlouthi

Software engineer and AI/ML engineer based in Tunis, Tunisia. I build practical
AI systems: voice agents, agentic platforms, and applied machine learning. I am
also a software engineering student at INSAT, graduating in 2027.

- Website: https://kacemmathlouthi.dev
- Email: kacem.mathlouthi@insat.ucar.tn
- GitHub: https://github.com/KacemMathlouthi
- LinkedIn: https://www.linkedin.com/in/kacem-mathlouthi/
- X: https://x.com/KacemMathlouthi
- Résumé (PDF): https://kacemmathlouthi.dev/kacem-mathlouthi-resume.pdf

## Current role

**Software engineer at [Callab AI](https://callab.ai) (YC P26).** I joined
before Y Combinator and went through the batch with the team
([batch profile](https://www.ycombinator.com/companies/callab-ai)). I build an
AI voice agents platform for on-prem telephony that works with existing PBX,
SBC, and contact center infrastructure, with no migration needed.

## Experience

**Google Summer of Code 2025 fellow at NRNB** (National Resource for Network
Biology).
[Project archive](https://summerofcode.withgoogle.com/archive/2025/projects/vFE5LPKW).
I built [VCell-AI](https://github.com/virtualcell/VCell-AI), an AI agent
platform that lets computational biology researchers query, explore, and
generate biomodels in natural language.

**Machine learning engineering intern at [Orange](https://www.orange.tn).** I
built an edge-optimized CNN for plant disease classification and an agentic
report generation pipeline.

## Projects

- **[Animus](https://tryanimus.app/)**: a platform that turns any topic into
  narrated, research-grounded Manim explainer videos.
- **[Metis](https://github.com/KacemMathlouthi/metis)**: an AI platform for
  GitHub pull requests, with a cloud coding agent for autonomous code review,
  issue resolution, and PR summaries.

## Research

Paper accepted at TMLR (Transactions on Machine Learning Research):
**"Generalization Measures under Controlled Covariate Shift: A Regime-Aware
Benchmark"** ([OpenReview](https://openreview.net/forum?id=X4RoujAYnY)),
co-authored with Sora Nakai (first author), Hiroki Naganuma, Youssef Fadhloun,
Kotaro Yoshida, and Ganesh Talluri.

The work asks whether generalization measures that look reliable under IID
evaluation still hold up when image classifiers face corruptions and
perturbations. It extends the benchmark of Jiang et al. (2020) to 40+ measures
on CIFAR-10-C/P, adding calibration and confidence measures and information
criteria, studied through rank correlation, local reliability, and
decision-level model selection. The central finding is that predictivity is
strongly regime-dependent: measures that work under IID evaluation do not
necessarily transfer under distribution shift.

## Education and community

Software engineering student at
**[INSAT](https://insat.rnu.tn)** (National Institute of Applied Sciences and
Technology) in Tunis, graduating in 2027.

Served as **Technical Manager** for
[AINS 3.0](https://www.ai-national-summit.tech), the AI National Summit
organized by [IEEE Computer Society INSAT Chapter](https://insat.ieee.tn/),
leading a three-track hackathon, a data science competition, and four
workshops.