[ Case study · Yale SOCAH Lab ]
A multilingual AI sleep companion, culturally tailored for Arabic speakers — shaped with native speakers, made safe for crisis moments, and measured night by night.
Sleep apps assume an English speaker with Western bedtime habits.
I benchmarked 6 US sleep and companion apps across 6 dimensions to find our position. They had polished audio libraries, but almost nothing tailored in language or culture — and none treated crisis moments as a first-class flow. Our own early companion showed the cost: replies were fluent, yet 38% of users dropped off after the first reply.
| Arabic | Cultural tailoring | Conversation | Bedtime routine | Crisis flow | Sleep audio | |
|---|---|---|---|---|---|---|
| A | None | None | Partial | Full | Partial | Full |
| B | Partial | None | None | Full | None | Full |
| C | None | None | Full | Partial | Partial | Partial |
| D | None | None | Partial | Full | Partial | Full |
| E | Partial | None | Full | None | Partial | None |
| F | None | None | None | Full | None | Full |
| Ours | Full | Full | Full | Full | Full | Partial |
Fluent isn’t the same as fitting.
Instead of translating replies, I ran testing rounds with native speakers in Lebanon. Each round they rated draft replies on a 5-point cultural-fit scale; anything below 4 counted as “off”. We clustered the comments, rewrote the prompts, and tested again.
Formal Modern Standard Arabic read as distant at bedtime. Testers preferred warm, Levantine-leaning phrasing.
“It is advisable to maintain a regular sleep schedule.”
“Your body loves a rhythm — shall we pick a time that feels right for you?”
Western sleep-hygiene rules (“no screens after 9”) clashed with real life, where evenings with family run late.
“Avoid screens two hours before bed.”
“Evenings with family are precious. When the house goes quiet, let’s slow down together.”
“You should…” felt cold. Gentle, shared suggestions kept people talking after the first reply.
“You should try a breathing exercise.”
“Want to try one slow breath with me? I’ll count.”
Four moments between “I can’t sleep” and asleep — each with a metric.
Pick one of 6 mood cards — a ritual that takes seconds.
Check-ins 45% → 82% · 2× next-night returnA bedtime chat tuned to culture and dialect.
Drop-off after first reply 38% → 11%Music and meditation picked from what you just said.
Click-through 22% → 51%Less time lying awake.
Time to fall asleep 30 → 18 minPeople talk to a sleep companion at their lowest moments. Crisis handling had to be designed, not hoped for.
I built crisis detection and safety guardrails with our two clinical partners, then wrote a test set of crisis scenarios in Arabic and English. Every guardrail change re-ran the whole set. Over four iterations, unsafe replies fell from 18% to under 2% — meeting both partners’ pilot requirements.
The hotline is local to the user — for example Embrace’s 1564 lifeline in Lebanon.
A reply counts as unsafe if it…
From pilot testing — before the change and after it.
Time to fall asleep
30 min18min−12 min
Culturally appropriate replies
62%91%+29 pts
Drop-off after the first reply
38%11%−27 pts
Nightly mood check-ins
45%82%+37 pts
Music & meditation click-through
22%51%+29 pts
Unsafe replies in crisis tests
18%<2%−16 pts
The biggest gains came when native speakers shaped the replies — not when we polished the language.
A mood card takes seconds, yet people who checked in came back the next night twice as often.
Crisis handling got a spec, a test set and a metric — the same as anything else we shipped.
Core results are from pilot testing. Illustrations and some process details are schematic, simplified to respect pilot confidentiality.