A proof-of-concept AI career coach grounded in real state data

Start with the honest framing. MyCareer.NJ is live for 1.7 million residents. The shared AI safety infrastructure behind this work runs in production too: the guardrails, the input checks, the output validation, and the kill switch that any AI feature would sit on top of. The career coach is the part that isn't there yet. It's a proof of concept, open for review and waiting on a pilot with the state, and no resident has used it. This case study is about how it was built to earn trust before it ever reaches one person.

Generative chat is powerful. In benefits and workforce settings it's also dangerous, because a confident wrong answer can send someone down the wrong path. Chelsea built the proof-of-concept coach so it can't do that. It answers only from real, current state training-program data, and when it doesn't have the information, it says so rather than guessing.

Built to be trustworthy

The system runs on Amazon Bedrock and pairs the model with retrieval over the state's real program data, so answers stay anchored to what actually exists. Testing for responsible AI ran in both English and Spanish, using adversarial and edge-case questions written to catch a wrong or invented answer before any resident could. A review across all fifty states found no comparable public, conversational tool grounded in real program data.

How grounding works here

• A resident's question would pull relevant, real program records first
• The model is constrained to answer only from those records
• When the data doesn't cover the question, it says so instead of guessing. It can't invent a program, a price, or a promise

The assistant on this site works the same way. It answers only from a small, curated knowledge base, and it won't make things up.