Most AI startups in 2024-2026 built one general-purpose assistant. We did the opposite. We built four specialized AI agents for the Gulf domestic-worker market — each for a different audience.
The 4 agents
| Agent | For | What it does |
|---|---|---|
| Hayat | Gulf families | Speaks Gulf-Arabic dialect, knows prayer times, plans halal/Ramadan menus |
| Nadia | Families hiring | Walks them through MOHRE/Musaned/PAM/MOL/LMRA verification |
| Layla | Recruitment agencies | Runs WhatsApp Business 24/7 in 7 languages |
| Amina | Migrant domestic workers | 9 free courses + daily wellbeing check-ins + abuse alerts |
The engineering decisions that matter
WhatsApp-first. No app. 99% of Gulf families and 3.5M Gulf domestic workers already use WhatsApp daily. Building a separate app would have added friction and reduced reach.
7 languages. Arabic Gulf-dialect, English, Tagalog, Hindi, Indonesian, Amharic, Swahili. Each language requires its own model fine-tuning, cultural adaptation, and content review.
Specialization beats generalization. A general assistant tuned for the worker would lose context for the family, and vice versa. The family asks "what should we cook for iftar?" — the worker asks "what should I do if I'm not paid this month?" Both are valid, both need specialized models and specialized safety guardrails.
Cultural fluency engineering. Prayer times, halal-by-default cooking rules, Ramadan iftar planning, 5-prayer scheduling — these are not nice-to-haves; they're table stakes for a Gulf product.
The result
Three of the agents (Hayat, Nadia, Amina) are completely free. The fourth (Layla, for agencies) is the only paid tier. This bias toward free works because the more people who use the agents, the more useful the underlying ecosystem becomes.
Hub for all 4 agents: https://www.gccdomestic.com/en/ai-agents/
Individual agent pages:
- Hayat (Gulf families): https://www.gccdomestic.com/en/hayat/
- Nadia (hiring guide): https://www.gccdomestic.com/en/nadia/
- Layla (agency operations): https://www.gccdomestic.com/en/layla/
- Amina (worker support): https://www.gccdomestic.com/en/amina/
Built by GCC Domestic — the verified directory of MOHRE/Musaned/PAM/MOL/LMRA-licensed recruitment agencies across the 6 Gulf countries. https://www.gccdomestic.com
What do you think about specialization vs generalization in vertical AI? Reply in comments.
Top comments (1)
Four specialized agents over one generalist is the right call and your reasons probably go deeper than they first look. A narrow agent isn't just better at its domain, it's dramatically easier to make reliable: a bounded job means you can actually define what correct looks like, test it, and constrain what it's allowed to do, whereas a do-everything assistant has a failure surface you can't fully enumerate, let alone QA. Specialization also unlocks cost control, each agent can run on the cheapest model that clears its specific bar instead of paying frontier prices for every interaction because one of them occasionally needs the big model. And it contains blast radius: if the family-facing agent misbehaves, it can't touch the worker-facing flow. The generalist's hidden tax is that every new capability you add raises the odds it does the wrong thing somewhere unrelated. The tradeoff you're now managing is the seams, routing the user to the right agent and handing off context between them cleanly, which is where a lot of multi-agent systems get messy. Narrow agents you can trust, composed carefully, beat one clever agent you can't fully predict. That bound-each-agent-so-you-can-verify-it instinct is core to how I think about Moonshift. How are you handling handoff when a user's need spans two of the four agents?