R-01
Arabic-First Language Systems
Dialect-aware models and Gulf Arabic benchmarks for a language the field still underserves.
Independent AI R&D Lab — Sultanate of Oman
Mebhath is an independent research and development lab in Oman. We investigate hard problems in Arabic-first language systems, autonomous agents, and applied machine intelligence — then engineer the findings into systems that hold up in production.
ب·ح·ثBAḤATHA — TO SEARCH DEEPLY, TO INVESTIGATE
Scroll
R — Research areas
Every area is held to the same bar: novel enough to publish, rigorous enough to run in production.
R-01
Dialect-aware models and Gulf Arabic benchmarks for a language the field still underserves.
R-02
Multi-agent orchestration, tool-use reliability, and long-horizon task planning.
R-03
Retrieval architectures, domain adaptation, and evaluation pipelines that survive contact with production.
R-04
Document intelligence and industrial inspection tuned for regional deployment conditions.
R-05
Arabic speech recognition and synthesis, and real-time voice agents across dialects.
R-06
Red-teaming, benchmark design, and governance frameworks for systems in the wild.
M — Method
A fixed pipeline, because rigor is a process property — not a personality trait.
01
Every project begins as a falsifiable question with explicit success criteria — not a feature request.
02
Small, fast experiments. We optimize for learning rate, not demo polish.
03
Benchmarks, evals, and adversarial testing before anything earns the word “works”.
04
Hardening, monitoring, and handover. Research that stays in a notebook is a draft.
L — Lab programs
Long-running programs where the research compounds.
P-01
An open evaluation suite for Arabic large language models.
In developmentP-02
An orchestration framework for reliable multi-agent systems.
InternalP-03
A document-intelligence pipeline built for Gulf enterprise conditions.
PilotA — The lab
Mebhath exists to close a specific gap: the distance between frontier research and what actually gets deployed in this region. We publish what we learn, build what we validate, and hold Arabic to the same engineering standard the field grants English.
If it only works on stage, it doesn't work.
Not a translation layer — a native constraint from day one.
We earn the right to grow a system by proving it small.
C — Contact
Research partnerships, pilot programs, and serious problems welcome.
hello@mehath.ai