Lab · AI on this site

Ask the Portfolio

A one-line field on the home page that answers questions about my work in plain text, with a link to the case study behind each answer.

Text onlyModel output is inserted as plain text, never as HTML
CitedEach answer ends with a From line linking the case studies it drew on
ScopedIt answers from my case studies and work history only, and says so when it can’t

Visitors have questions the page order can’t answer

Hiring managers ask about pushback, failure and influence. A one-line field answers those directly, and its answers took the place of the “How I work” and “Featured case studies” sections on the home page.

The home-page ask field with a sample question, an AI answer panel and source chips pointing to case studiesThe home-page ask field with a sample question, an AI answer panel and source chips pointing to case studies

What I did

I wrote a private knowledge file from my case studies and work history, in my voice, together with the rules the assistant must follow.

A serverless function sends each question, plus the last few exchanges so follow-ups make sense, to a language model along with that file, and streams the answer back into the page.

How it works

1

Ask

Type a question or pick a suggested one. Pressing “/” jumps to the field from anywhere on the page.

2

Answer from one source

The function answers from the private knowledge file. The rules: use numbers exactly as written, never invent a project, and never claim credit that belongs to a team or a program.

3

Show the work

The answer streams in as text and ends with a From line that links to the case studies behind it.

4

Know when to stop

A rate limit protects the endpoint. The panel reminds visitors that AI can be wrong and links to my email for anything important.

Decisions that matter

✓

Small counts stay counts: “50 invited, 11 responded,” never a percentage of a small number.

✓

Program and company results belong to the program and the company, not to me, and the assistant is told so.

✓

It declines or redirects questions about compensation, personal details and personality profiles, and anything the file doesn’t cover.

✓

Before launch I wrote a 40-question test set: behavioral questions, numbers that must match, attribution traps, questions it should decline, and attempts to break it.

✓

The knowledge file stays on the server and is not publicly served.

Honest limits

It is still a language model, so it can get things wrong, which is why the panel says so and points to my email. Unlike the scripted AI in my patient-portal prototype, this one is real, and that is why it needed the rules above.

Want to ask it something harder?