I build systems that run unattended, and I measure what they inherit. Everything runs on-device — a phone, a laptop, no rented GPUs. That constraint is the design, not the compromise.
Website · marwan-naili.me — research record — method paper
[A] Agentic systems Runtime containment, capability security, the parts that only fail under real load.
- Minis-X — agent runtime for Android: on-device model routing, tool execution, Linux sandbox for untrusted code, plugin permissions with a network allowlist enforced below the app layer, encrypted vault, run telemetry. Signed CI builds, verified on-device after every release.
[B] Applied ML & evaluation Retrieval that is allowed to say I don't know, and evals that don't flatter me.
- CyberRAG — retrieval arsenal over WSTG / ATT&CK / field notes. Keyed CVE lookup and concept corpora in separate indexes so a 1,900-record catalogue cannot poison concept queries. Abstention is a valid output.
[C] Security measurement Passive-first. No exploitation, no credential testing, no state-affecting action.
- DarkTortilla-RAT — reverse engineering of a commodity Android RAT's Telegram exfiltration channel. Samples in isolation; no live infrastructure contacted.
- algeria-wilayas — 69 wilayas + communes, 2026 reform, reproducible build.
Observed / reported / inferred / unknown are kept apart. A passing command proves the mechanism worked, never that the objective was met.