Community IVR: Voice AI for offline communities (Cal Hacks 8.0)
During the COVID-19 pandemic, Indonesian government health guidance, social program eligibility, and emergency services information was primarily distributed through smartphone apps and social media platforms. This created a critical information exclusion for the estimated 40% of Indonesians without smartphones -- predominantly rural, elderly, and low-income populations who were simultaneously the most vulnerable to COVID-19 health and economic shocks.
IVR systems fail in emerging markets through predictable usability breakdowns: menu trees designed for literate, patient users collapse under dialect variation, background noise, and low digital confidence. Indonesian ASR trained on formal broadcast speech systematically underperforms on Javanese-accented, Sundanese-accented, and code-switched Indonesian -- the dominant registers of the target population. The knowledge base itself was a moving target: quarantine regulations, bantuan sosial eligibility criteria, and emergency contact routing changed weekly during the 2020 pandemic surge. A static knowledge graph would be obsolete before deployment.
Designed a phone-based AI assistant operable via basic feature phones on 2G networks -- no smartphone, no internet, no app required. Core components: ASR (Automatic Speech Recognition) optimized for Indonesian language and local dialect variation (Javanese-accented Indonesian, Sundanese-accented Indonesian), a structured knowledge graph for COVID-19 health guidance, government service eligibility, and emergency contact routing, and a Text-to-Speech response layer.
Designed for zero digital literacy barrier: interaction model mirrors existing DTMF IVR systems that Indonesian users were already familiar with from banking and telecom services. Free-form voice queries supported for users unfamiliar with menu navigation.
Knowledge graph manually curated for Indonesian context: local health authority contact numbers, regional quarantine regulations, bantuan sosial (social assistance) eligibility criteria, and local dialect variation in query phrasing.
Directly addressed digital exclusion at a population scale -- the 40% of Indonesians without smartphones are invisible to virtually all AI applications. Won Best Community Track at UC Berkeley's flagship hackathon, validating the design approach with an international jury. The architecture pattern (feature phone + ASR + knowledge graph, no smartphone required) remains relevant for digital inclusion challenges across Sub-Saharan Africa, rural South Asia, and other underserved markets.