Personal site / September 2026
Builder. Thinker. Doer. Researcher.
I build software and study how intelligent systems make decisions.
Currently building Aurels and researching reliability in AI agents.
Casablanca, Morocco
Selected workThree pieces of work. The rest is an archive of experiments and useful detours.
A control layer for autonomous AI agents. Aurels evaluates sensitive actions before execution, applying policy and semantic risk checks before an agent can perform irreversible operations.
Co-founded Preparly, a gamified exam-preparation platform, and served as CTO / Product Designer. I worked across product, UX and engineering, from early prototypes to production infrastructure.
Reached approximately 5,000 monthly active users at peak.
Selective Risk Control for LLM Decision Agents under Uncertainty studies whether binding adversarial critique and abstention gating improve decision behavior when evidence is sparse or contradictory.
The result is intentionally narrow: the mechanism acts primarily as a conservative risk gate. It reduces exposure to difficult cases through abstention, but does not consistently improve active-decision performance at matched coverage.
Aurels is not a model. It is a decision layer around a model’s ability to act.
Research is a way to make the limits visible before systems are asked to do more.
Manuscript under external peer review
The work evaluates a controlled sparse-and-contradictory evidence benchmark: five agent architectures are asked to decide, defer, or reject. It tests whether critique changes the quality of decisions—or simply changes when a system refuses to make one.
Main finding: binding critique behaves as a conservative risk gate. It can be useful when false commitment is costly and abstention is acceptable; it is not a universal upgrade to active decision-making.
The paper’s limits matter: the evaluation uses a controlled benchmark and proxy regret, so its empirical result is domain-bound rather than a claim of universal safety.
Things I built to learn from. Not every prototype needs to become a company.
An experiment in intentional, non-addictive social software.
An AI venture-building experiment. Kept as an exploration, not a main company.
A sustainable-travel project built with a student team.
Curiosity is more useful when it becomes shared infrastructure.
An entrepreneurship initiative and community involving approximately 20 students.
A leadership role in a ~30-student research community, with work around IoT and AI.
Technology communication, public speaking and small explanations are part of the practice—not a separate personal brand.
A few working principles, revised whenever the work proves them wrong.
I learn fastest when an idea becomes something people can actually use.
A system becomes interesting when it leaves the controlled environment and starts making decisions in the real world.
In products, research and design, knowing when not to act can matter as much as knowing what to do.
The through-line is less a career ladder than a sequence of things that taught me what to try next.
Co-founded Aurels and began building an independent control layer for sensitive AI-agent actions.
Completed independent research on selective risk control for LLM decision agents. The manuscript entered external peer review at Discover Artificial Intelligence.
Transitioned from Preparly after the platform reached approximately 5,000 monthly active users at peak.
Co-founded Preparly and built the product as CTO / Product Designer.
Expanded involvement in research and student entrepreneurship.
Worked on explainability and reliability questions in machine learning.
Built an early experiment around intentional, non-addictive social software.
Built an early sustainable-travel product with a student team.
Software, design, automation and digital products—mostly a way of finding problems worth staying with.
Earlier short-form work across UX/UI, marketing, branding, automation, frontend development and video editing helped build range. It is kept here as context, not as the main story.
Outside the terminal
Long-distance running and triathlon training; reading about technology, economics and systems; design; and building unnecessary side projects because I am curious.