How the assistant on this site works

You ask about my work, and it answers in a few short sentences. If the answer is about one project, the page behind it moves to the section that proves it, like the results or the edge cases. If it covers a few projects, you get a card for each one under the answer.

On a phone the chat covers the screen, so instead of moving the page it gives you a button to go there.

An AI that makes things up about you is worse than no AI. So it can only answer from one knowledge file: my case studies, my resume and my LinkedIn, which I read through and approved. It doesn't search the web.

When it doesn't know something, it says so and gives my email. Numbers have to match the case studies exactly. Knowt was never tested with students, so it won't invent results for it. Salary, visa and other personal questions go to my email too.

If someone asks for a poem or tries to make it ignore its rules, it gives one fixed line: "I'm only here to talk about Manoj's work."

The model has one tool: show a case study, optionally at a section. It calls that before it writes, so the page moves while you read. If it forgets, a backup reads your question for a project name and a word like "results" and moves the page anyway.

Every move comes with a "Go back" link that puts you where you were.

I gave it the same rules I write by: short plain sentences, no "Great question!", none of the usual AI words, and no offer to help at the end. It writes plain text, and the chat strips any formatting marks that slip through.

It works with just a keyboard. Esc closes it and puts focus back where you were. Screen readers hear each answer when it finishes, and hear where the page moved. On phones it covers the screen and keeps focus inside until you close it. Errors say what happened and how to reach me, instead of a vague "something went wrong".

Before launch I asked it the questions I was worried about. It refused to share my phone number, sent salary questions to my email, said no when asked if Knowt improved retention, and gave its one fixed line when asked for a poem. I also checked that factual answers matched the case studies and that the page landed on the right section.

Testing it in a real browser also found bugs that unit tests missed: follow-up questions failing, answers repeating, and the page stopping short of the section. Those are all fixed.

Voice mode, so you can ask out loud. A proper set of test questions that runs every time I change the knowledge file. And a look at which questions it couldn't answer, so I know what my case studies are missing.