KKaebi: Voice-First Companion AI

Inspiration

KKaebi initially breathed life as "깨비", an on-device local voice assistant tailored for Helios laptops, handling offline commands, riddles, and simple quizzes. However, as we observed the rapid aging of society and the digital exclusion of children, we asked ourselves a fundamental question: How can we liberate AI from heavy screens and expensive hardware to make it truly ubiquitous?

We found our answer in wearables. By separating KKaebi from desktop hardware and pivoting it into a cloud-native, voice-first companion AI, we realized we could deliver direct value to those who need it most. KKaebi was reborn to solve two critical social challenges: accelerating early language development in children through fun interactive edutainment, and maintaining cognitive health (dementia prevention) in seniors through continuous mental stimulation.

What it does

KKaebi is a screen-free, voice-only companion that provides seamless cognitive activation and engaging learning experiences directly on compact wearable devices (e.g., smartwatches, screenless clip-on bands).

Dynamic Riddle & Quiz Playground: Offers context-aware, bidirectional verbal riddles that encourage children to think critically and speak expressively.

Adaptive Cognitive Exercise (AI Problem Bank): Generates personalized quiz sessions. The system dynamically adjusts the vocabulary level and complexity of questions based on the user's verbal reasoning speed and accuracy.

Zero-Touch UX: Completely eliminates the need for physical typing or navigation, breaking down accessibility barriers for elderly users with visual impairments or motor difficulties.

How we built it

To guarantee a natural, human-like flow on low-power devices, KKaebi was built with an "AI-Native" architecture centered around the Gemini 2.5 Flash model and the Antigravity framework.

Traditionally, voice systems run a multi-stage cascade pipeline consisting of Speech-to-Text (STT), Large Language Model (LLM) processing, and Text-to-Speech (TTS). We modeled the latency of this traditional pipeline as:

$$T_{\text{cascade}} = T_{\text{STT}} + T_{\text{LLM}} + T_{\text{TTS}} + T_{\text{network_overhead}}$$

Because $T_{\text{cascade}}$ often exceeded $2.5\text{s}$ due to serialization overhead, it ruined conversational spontaneity on resource-constrained wearables. To solve this, KKaebi leverages Gemini's native audio-to-audio streaming endpoints. Under this model, the processing latency is simplified to:

$$T_{\text{native}} = T_{\text{audio2audio}} + T_{\text{network_stream}}$$

By bypassing the intermediate text extraction and synthesis stages, we achieved an ultra-low response latency of $T_{\text{native}} < 0.8\text{s}$. The Antigravity framework acts as our orchestration layer, tracking user states, computing cognitive performance metrics, and managing session-by-session context memory without overloading the wearable's physical hardware.

Challenges we ran into

Our biggest roadblock was the hardware-level physical constraints of wearables. Smartwatches and smartbands have tight CPU limits, extremely restricted memory, and very small battery capacities. Running heavy computational tasks locally on the device was out of the question.

Furthermore, because these devices are used on the move (outdoors, playgrounds, public spaces), handling heavy ambient noise while keeping the streaming connection stable on fluctuating mobile networks was a significant challenge. We solved this by implementing an adaptive WebRTC audio gateway. It dynamically compresses and packetizes voice data at the edge before streaming it to Gemini, allowing us to maintain robust voice connectivity and sub-second latency even in suboptimal network conditions.

Accomplishments that we're proud of

We are incredibly proud of migrating a hardware-dependent, legacy laptop voice engine into an agile, highly accessible SaaS API model. We proved that true generative AI can operate fluently on a screenless wristband with zero typing barrier.

Creating an inclusive design that brings joy to toddlers who cannot write yet, while simultaneously serving as a gentle, warm companion for elderlies experiencing cognitive decline, is our greatest triumph. We successfully transformed the concept of "AI Companion" from a sci-fi dream into a practical, daily tool.

What we learned

We learned that user experience in voice-first AI is entirely defined by latency and listening behavior. Even a $1.5\text{s}$ pause can make users feel disconnected or prompt them to interrupt the AI. Optimizing native audio channels was not just a technical choice, but a requirement for psychological comfort.

Additionally, we realized the immense power of context tracking. Standard RAG (Retrieval-Augmented Generation) was insufficient for cognitive training; we had to develop a structured logic that tracks not just what was said, but the cognitive friction—the delay in user answers—to dynamically adapt the complexity of our interactive quizzes.

What's next for KKaebi: Voice-First Companion AI

Our immediate goal is to launch a Freemium mobile/wearable application, offering daily free cognitive quizzes while unlocking unlimited customized learning modules and deep cognitive health analytics for $4.99/month.

To drive scalable, real business revenue, we are pursuing strategic B2B partnerships with local senior care centers, smart-toy creators, and wearable manufacturers to pre-integrate the KKaebi API into kids' and seniors' smartbands and wearable devices worldwide.

Built With

• and-content-curation-dashboards.-databases:-a-dual-db-strategy-is-utilized.-mysql-handles-rapid-web-application-transactions
• and-metadata.-oracle-database-is-deployed-to-securely-manage-and-run-complex-analytics-on-highly-structured
• ciboard
• comforting
• enterprise-grade-hybrid-architecture-designed-to-bridge-legacy-stability-with-cutting-edge-ai-real-time-capabilities:-frontend-&-web-admin:-built-using-php-and-the-ciboard-(codeigniter-based)-framework
• linux
• long-term-user-cognitive-training-logs-and-enterprise-b2b-partner-data.-ai-orchestration-&-reasoning:-powered-by-gemini-2.5-flash-for-lightning-fast-cognitive-processing
• melotts
• mysql
• onnx-optimized-execution-on-edge/on-device-environments
• oracle
• orchestrated-via-the-antigravity-framework-to-manage-session-by-session-state.-voice-synthesis-(tts)-pipeline:-we-leverage-the-state-of-the-art-supertone-voice-api-(powered-by-the-supertonic-engine).-by-utilizing-supertonic?s-ultra-fast
• php
• running-securely-on-linux-(ubuntu/centos)-servers.-this-setup-manages-user-licenses
• user-sessions
• we-deliver-deeply-emotive
• web-based-subscription-portals