RESEARCHER — AGENTIC SYSTEMS / APPLIED ML / SECURITY
ALGIERS, DZ · UTC+01:00
I build systems that run unattended, and I measure what they inherit. Three areas: agent platforms where a model calls tools and can break its own environment, applied ML where the hard part is knowing when a system is wrong, and security measurement — exposure, malware, and the surface area between them.
Everything runs on-device. A phone, a laptop, no rented GPUs, no cloud dependency. That constraint is not a limitation I work around; it is the reason the systems still function when the network doesn't.
01 — PILLARS
Platforms where a model calls tools, executes code, and runs for hours without supervision. I work on the parts that only fail under real load: sandbox containment, plugin permission models, network allowlisting enforced below the application layer, loop detection, and the CI discipline to ship a fork for years without upstream drift.
Retrieval that is allowed to say I don't know, and evaluations that don't flatter me. Supervised fine-tuning on constrained hardware, routing across small specialist models, adversarial needle-retrieval against long context, and building a gold set by hand so a benchmark means something.
Exposure enumeration, malware and exfiltration analysis, attack-surface review of real applications. Passive indexes, protocol identification, static and dynamic analysis on samples already in hand. No exploitation, no credential testing, no state-affecting action — the line is cheaper to hold before you cross it.
02 — SYSTEMS
verified 2026-10-01
A complete agent runtime for Android: on-device model routing, tool execution, a Linux sandbox for untrusted code, plugins with declared permissions and a network allowlist, scheduled tasks, an encrypted vault, and run telemetry. Ships through CI as a signed build and is verified on-device after every release — because a green build is a claim, not a test.
verified 2026-09-28
Entity intelligence that runs entirely from a phone. Content-addressed evidence in SQLite, a resolver that returns SAME / DIFFERENT / UNKNOWN with the evidence attached, per-source coverage states so "not checked" can never be mistaken for "nothing found", and report generation behind a citation validator that fails closed.
verified 2026-09-29
Reverse engineering of a commodity Android RAT's Telegram exfiltration channel: payload extraction, protocol reconstruction from recovered code, derived indicators of compromise. Samples handled in isolation; no live infrastructure contacted.
updated 2026-09-28
A cron-snapshot geospatial feed on edge workers — earthquakes, military and civil flights, satellites, fires, hazards, news — collected on schedule into a historical store. Deliberately inverted from per-request fan-out: the observation is the record, not a query. A national Algerian-scoped variant runs alongside it.
updated 2026-09-29
A retrieval arsenal over WSTG, ATT&CK, exploitation references and my own field notes, rebuilt around one rule: abstention is a valid output. Keyed lookup data and conceptual corpora live in separate indexes so a 1,900-record CVE catalogue cannot poison concept queries, and a query term absent from the corpus is treated as evidence the corpus lacks the topic.
started 2026-09-28
An unanswered question: what is the real capability floor of small open-weight models used as autonomous cyber tooling? Existing public claims test one deliberately undisclosed model size and call it a result. I'm building the sweep — size against task fitness, below single-GPU scale, micro-policies versus language models, payload execution versus rhetoric.
verified 2026-09-28
All 69 wilayas and their communes, updated for the 2026 administrative reform, bilingual naming, with a documented generation procedure — a reproducible build rather than a one-off export.
03 — METHOD
04 — IN PROGRESS
started 2026-09-30
Passive-first measurement of Algerian critical-sector exposure — energy, water, ports, rail, telecom edge. The deliverable is a chain, not a port number: observation → attribution → technology → known vulnerability → how long it has been observable. Rules of engagement published before collection; disclosure to the responsible party and the regulator first.
started 2026-09-28
Model size swept against autonomous cyber-task fitness below single-GPU scale, establishing the quantitative floor that current public claims only gesture at.
05 — CONTACT
SECURITY — security@marwan-naili.me
CODE — github.com/hellbound2307
RESEARCH RECORD — /.well-known/security.txt
DISCLOSURE TERMS
Security acknowledged within 3 business days. Publication after 90 days of silence, unless fixed.