# Vrushabh Jambhulkar

> Senior Software Engineer — AI Product Engineering

Senior Software Engineer at Coinbase focused on building AI-native products, platform infrastructure, and developer tooling at scale. Over the past several years, evolved from mobile and web application development into platform architecture, AI-powered workflows, and cross-functional technical leadership.

Work spans frontend systems, backend-integrated platforms, agent-based workflows, MCP integrations, evaluation frameworks, experimentation systems, and large-scale product experiences used across Coinbase. Most recently, architected and helped build a company-wide content operations platform that combines schema-driven UI generation, AI agents, tool orchestration, analytics, experimentation, and content management into a unified workflow.

## Site

- [Full profile as markdown](https://vrushabhjambhulkar.com/index.md)
- [Resume (experience, skills, education — no projects)](https://vrushabhjambhulkar.com/resume.md)
- [Experience](https://vrushabhjambhulkar.com/experience.md)
- [Skills](https://vrushabhjambhulkar.com/skills.md)
- [Education](https://vrushabhjambhulkar.com/education.md)

## Projects

- [Ember — AI companion pendant](https://vrushabhjambhulkar.com/projects/ember.md): A wearable pendant (ESP32-S3 circular AMOLED) with camera, mic, and speaker, driven by parametric commands from a cloud backend — an expressive animated face with emotional presence.
- [Stocker — Ambient stock & crypto dashboard](https://vrushabhjambhulkar.com/projects/stocker.md): A single-owner ambient dashboard that cycles through full-screen slides — price charts, sector heatmaps, market movers, MACD signals — built to run 24/7 on an always-on display with server-side caching and multi-provider fallback.
- [AI Product Engineer Copilot — A multi-step agent that plans, cites its sources, and grades its own work](https://vrushabhjambhulkar.com/projects/ai-product-engineer-copilot.md): A LangGraph multi-agent system that generates PRDs, user stories, experiment designs, architecture reviews, and roadmaps — a supervisor routes into a PRD agent, three sub-agents fan out in parallel once the PRD exists, and a roadmap agent joins their outputs. Grounded in two real MCP servers (pgvector docs search, live GitHub stats), with rate limiting, durable run traces, and a bias-checked eval harness.

## Blog

- [Your Agent's Output Was Valid. It Was Still Wrong.](https://vrushabhjambhulkar.com/blog/agent-output-valid-still-wrong.md): A production agent leaked a promotion to users outside its target group with every eval green. What that taught me about the gap between a valid artifact and a correct outcome — and the three layers of containment we built around it.