Goldi: a production AI career coach for 100K+ users

The problem

Students looking for internships and first jobs need coaching that is personal and available whenever they are, at a scale no human team can staff. The coaching also has to meet students where they already are: on the web, on their phones and inside the tools they use to job hunt.

Approach

• Built Goldi as a production career-coaching agent on Anthropic Claude, OpenAI and LangChain, with custom tool calling and multi-turn conversation handling.
• Added a voice mode with Hume AI, with natural interruption handling and tone-adaptive pacing.
• Built a WhatsApp and SMS consent and onboarding flow, plus real-time data pipelines for SMS nudges and behavior analytics.
• Built a Chrome extension (MV3 side panel, content script and auth bridge) that brings the agent into LinkedIn.

Architecture

• Next.js and TypeScript app with Clerk for auth and Twilio for messaging, built as a distributed onboarding system.
• Serverless backend on AWS Lambda with PostgreSQL and Drizzle ORM.
• Multi-playlist navigation with URL state, race-condition fixes and caching so the experience holds up at scale.
• CI/CD with automated testing, linting and environment-specific deploys to Vercel; Sentry for errors and PostHog for product analytics.

Results

• Voice mode shipped to 100K+ users through a Walmart partnership.
• The onboarding system scaled to 100K+ users with a 35% lift in verified opt-ins.
• SMS nudges and the WhatsApp and SMS flow drove a 28% increase in re-engagement.
• The move to serverless cut cold-start latency by 50%.

Stack

• Anthropic Claude
• OpenAI
• LangChain
• Hume AI
• Next.js
• TypeScript
• Twilio
• AWS Lambda
• PostgreSQL