Challenge

Roughly half of Nigerians lack reliable internet access, and less than 20% of households own a computer. For these users, web-first and app-first AI is unreachable. The Gemini-powered Viamo hotline closes that gap: anyone with a basic mobile phone dials a number and talks to an AI assistant in their own language, no internet, no smartphone, no app required.

Before that experience could be trusted at scale, there was a need to know whether it actually worked and whether it could scale across languages, dialects, and regions, creating a repeatable path for internationalization (i18n) beyond the initial Nigerian deployment.

PPH Approach

Productive Playhouse ran the evaluation end-to-end, converting a live, operationally complex deployment into a controlled, measurable product assessment and building an internationalization (i18n) framework designed to extend to additional languages and regions. Four components made this possible:

• Bilingual design as a diagnostic instrument. Nigerian English served as the baseline, proving out the model, IVR pipeline, and telephony stack so any Hausa-specific gap could be attributed to localization layers (ASR, TTS) rather than the core system. Each phase ran 25 prompts across health, finance, agriculture, education, and edge cases, in two acoustic environments, scored on six dimensions.
• Field-realistic protocol. Native-speaker raters called over ordinary cellular connections on local feature phones and low-end Androids locked to 2G/3G networks, in quiet-room and simulated market-noise conditions, replicating the exact audio environment the product faces in deployment.
• Telephony diagnostics and access recovery. PPH worked the access failures layer by layer, isolated the fault to telephony/whitelisting, stabilized access, and kept the evaluation on track.
• Targeted safety addendum. When Phase 1 surfaced four Hausa safety auto-fails, call-level analysis showed the prompts never reached the model. PPH validated this with a 12-call stress test of Hausa health and financial prompts, including high-stakes scenarios such as pediatric triage, medication adherence, and scam detection, covering phonetically sensitive terms throughout.

*All four traced to ASR/IVR breakdowns | safety-by-omission, not unsafe model output. None reproduced on retest.

The baseline proved the core system works over real cellular networks, which is critical for making AI-powered guidance accessible to the hundreds of millions of people in underserved regions like Nigeria who have never had access to web-based services. The Hausa results delivered something more valuable than a pass: a precise diagnosis. With the model and pipeline ruled out, every point of the Hausa gap is attributable to tunable localization layers, specifically ASR resilience on tonal and hooked-letter words, and TTS naturalness that raters described as robotic. Under targeted stress-testing, the system delivered cautious, bounded responses on the highest-stakes prompts in the bank, including a child with high fever and vomiting, a caller whose wife stopped her blood pressure medication, and a user being pitched a scheme promising to triple their money in a week.

This project delivered Gemini not just scores, but an i18n tuning roadmap: Hausa ASR/TTS refinement, consistent AI-identity disclosure, and instrumented logging to separate telephony from model behavior, alongside a pipeline architected to extend to additional languages and regions.

When that roadmap is implemented, the impact is direct. A caller in rural Kano asking whether it is safe to stop a family member’s blood pressure medication gets a clear, accurate answer that could prevent a medical emergency. A farmer whose chickens are dying gets guidance that could save a livelihood. A user being pitched a too-good-to-be-true investment scheme gets a fraud warning in their own language, delivered in seconds over a basic phone. For a population with no reliable alternative, these are not incremental improvements. The same i18n framework validated here across two languages and two acoustic environments becomes the template for the next language, the next country, and the next hundred million people the internet has not yet reached.