Bilingual AI Safety Companion
Cloning a domain expert's personality into a trusted, bilingual AI safety companion.
- Role
- Conversational AI Designer & Engineer
- When
- Sep 2025 – Feb 2026
- Context
- Honda Motor Co.
- Built with
- Figma, Gemini API, Neo4j (Graph+Vector RAG), FastAPI, React 19, Qwen3-TTS, MuseTalk
Overview
A bilingual (JP/EN) conversational AI system that models the personality of a real domain expert to deliver trusted cycling-safety guidance in their authentic voice. The core challenge: Japan's cycling regulations are dense and inaccessible — the existing resources are dry and intimidating. The solution makes compliance feel like talking to a knowledgeable friend, not reading a rulebook.
The problem
Japan's bicycle safety regulations are scattered across dense legal documents. Most cyclists don't know the rules. Existing educational resources are formal and off-putting. The design challenge: how do you make regulatory content engaging and trustworthy enough that people actually absorb it?
Decisions and status
Decision card
The call that changed the answers users got.
Where it stands
A prototype, described as one.
- PrototypeBilingual JP/EN assistant with a lip-synced voice avatar, escalated to General Manager level with a positive review.
- 54+Personality traits modeled
- 2Languages (JP/EN)
- GraphRAG architecture
- VoiceCloning integrated
Research and discovery
- Multimodal personality analysis — video transcripts, speaking patterns, humor markers, linguistic fillers
- Linguistic extraction — sentence endings, discourse markers, characteristic JP/EN code-switching patterns
- Regulatory document structuring — chunking cycling rulebook for accurate, citable RAG retrieval
- Conversational UX — balancing authentic personality with factual accuracy and citation requirements
Key insight
Personality isn't just tone — it's trust. Users engaged more deeply and retained information better when the AI felt like a real person with opinions, warmth, and humor rather than a neutral information retrieval system.
Design process
Personality modeling — extracting 54+ behavioral and linguistic traits via multimodal Gemini analysis
Prompt engineering — encoding fillers, humor style, sentence length, warmth markers into system prompts
Initial RAG — FAISS vector search for regulatory retrieval (returned correct but shallow answers)
Migration to Graph+Vector RAG (Neo4j) — connected regulations surface together in one coherent response
Voice avatar integration — MuseTalk for animated face lip-synced to Qwen3-TTS voice cloning output
Frontend — React 19 component library within Honda design system constraints
The pivot
FAISS-only retrieval returned isolated regulation fragments — when a rule connected to three others, users got one. Migrating to Neo4j Graph+Vector RAG let the system traverse the knowledge graph and respond with full regulatory context in a single natural response. This was the turning point from 'accurate but choppy' to 'genuinely feels like an expert.'
Results
- Authentic personality validated through user testing — participants reported feeling they were talking to a real expert
- Bilingual responses with natural Japanese discourse markers and appropriate cultural humor
- Source-cited answers grounding every factual claim in specific regulation articles
- Full voice avatar prototype with lip-synced animation received GM stakeholder approval
Reflection
The hardest design challenge wasn't the RAG pipeline or the voice cloning — it was crafting a system where every filler word, every joke, every moment of hesitation was a deliberate design decision that built trust. The engineering was in service of humanity.
- Conversational UX
- Personality Design
- Graph RAG
- Bilingual AI
- Voice Cloning