Local language
“Nipe statement yangu”
must keep working for
your English bot.
A language layer for desks that code-switch, tuned evals, locale prompts, and human-review loops for Kenyan contact centers.
Kenyan customers skip “language setting.” They mix English, Swahili, and Sheng in one WhatsApp bubble. Generic multilingual APIs trained abroad stumble on “Niaje, nataka ku-check balance yangu”, and stumble loudly in public.
Sample intents we tune
Real phrases, real phrases, off textbook drills.
customer "Nipe statement ya last month pls" bot [grounded] Ombi la taarifa ya mwezi uliopita… customer "Bana hii charge ni nini?" bot [retrieve fees KB] · cite · offer agent customer "Nataka kuongea na mtu" bot [escalate] human queue · thread attached
Deploy on WhatsApp desk, agent-assist sidecar, or internal training for new hires in Eldoret and Mombasa sites. Pair with red team, jailbreaks in Sheng are a category of their own.
FAQ
CX and language quality.
Is this machine translation bolted on?
No, retrieval, system prompts, and eval suites are built for code-switching and Nairobi street register, Nairobi street register, beyond dictionary Swahili.
Voice or text?
Both paths: WhatsApp/chat first; voice via transcript assist for agents (voice assist for agents, short of a synthetic call centre).
Will it invent tariffs or rates?
Grounded mode cites your KB. Off-domain questions escalate, we tune refusal behaviour in eval.
Can we add mother tongues later?
Pack architecture supports additional locales; priority is Swahili/Sheng/English mix for Kenyan CX.
Tune a desk for Kenyan speech
Send 50 real (redacted) chat logs. We benchmark baseline vs language pack on your stack.
Request language pilot