How it was built
A small product with the guardrails a school would ask for.
Vrymdor Pathway Coach was built on the Vrymdor engine, the same approach the Vrymdor platform uses for its own training simulations: a simulated conversation, facts locked to a source, and scoring that teaches. It was built AI native: designed and directed by its builder, with Claude Code writing and testing the code in the terminal.
Step 1
The facts come first
A Python script reads 32 career pages from the Bureau of Labor Statistics Occupational Outlook Handbook (public domain) and turns each into a fact card: pay, education, what the work is, the work environment, how to get there, and the job outlook. Each card keeps its source and the date it was read. Tables and chart labels are left out, so stray numbers never become "facts".
Step 2
A Mentor that only knows the card
Each Vrymdor Mentor is Claude, given one fact card and a role: answer as a working professional, at the student’s grade level, in under 110 words, and say "I’m not sure" rather than guess. Mentors say plainly that they are AI characters.
Step 3
A number gate in code
Instructions are not enough, so every reply is checked in code. Any figure that is not on the career’s card (a made up salary, a wrong percentage) sends the reply back once with a note; a second miss drops that sentence. The feedback card is checked the same way.
Step 4
Privacy by design
No accounts, no cookies, no analytics, no database. The browser holds the conversation; the server strips emails, phone numbers, addresses and names from every student message before a model sees it, and keeps nothing. Each Mentor speaks with its own ElevenLabs voice, cast to fit the character; only the Mentor’s own words are sent to be voiced, signed on the server so nothing else can be, and never what the student typed. There is no microphone input, because browser speech recognition can send audio to a third party.
Step 5
Feedback that teaches
When the student finishes, a second Claude call scores four interview skills (curiosity, preparation, clear questions, follow up) from 1 to 4, quotes the student as evidence, and writes a printable "What I learned" card from the fact card only. The answer is structured output, so the page always gets the shape it expects.
Step 6
Limits for a public demo
Per visitor and per day caps on model calls, ten questions per conversation, and refusal fallbacks on the server so a declined request is retried rather than shown as an error.
The stack
- Python (standard library only) for the fact builder, with unit tests on a saved page
- Next.js and React with TypeScript for the pages and the two API routes
- The Anthropic SDK: Claude for the Mentor (prompt cached per conversation) and structured output for the feedback
- Zod to validate every request and every model answer
- Tests for the privacy filter and the number gate, run with every change
What a district version would add
- Careers mapped to a district's own pathways and local employers
- Rubric scores aligned to a district's Portrait of a Graduate
- Teacher view of class level, anonymous trends, with no student identities
- Spanish and other languages, and a review of every Mentor by a real professional