Roberto Yamanaka

• Shipped a real-time voice AI language coach end-to-end solo: React Native mobile app (iOS/Android), Next.js 16 full-stack portal with auth and subscriptions, and a Python LiveKit agent running a streaming STT → LLM → TTS pipeline over WebRTC.
• Designed an automated benchmarking system that evaluates conversation quality across four dimensions (vocabulary, grammar, pronunciation, fluency) using Gemini 2.5 with structured JSON outputs.
• Built data pipelines for vocabulary ingestion, Japanese morphological analysis (Fugashi / MeCab), and spaced-repetition sync between mobile SQLite and Postgres via Prisma.
• Self-hosted the production ML serving stack (Whisper STT, GPT-4.1, Hume / Inworld TTS) with Docker, health checks, Sentry tracing, and VPN-only SSH access, operating at sub-second end-to-end turn latency.

App Store ↗

• Built an MCP server enabling AI agents to interact with Starknet wallets through natural language, directly applicable to autonomous agent workflows.
• Architected a composable TypeScript SDK for wallet infrastructure, adopted by 150+ startups.

Website ↗

• Led a team of 5 engineers to build a digital identity wallet for Mexican NeoBank Banco Sancus, from requirements to production.
• Defined technical architecture, sprint cadence, and code standards across the full product lifecycle.

• Built multimodal ML pipelines (OCR, Whisper transcription, NLP extraction) that ingested unstructured qualitative research and surfaced structured insights through full-stack dashboards used by research teams — bridging research output and product UI.

• Designed serverless pipelines on AWS + Databricks processing 100k+ government documents and millions of e-commerce SKUs for inflation prediction models.

Website ↗