shivam.
Open to roles
View work Résumé
← Back to the portfolio

A working product experiment

A model makes a judgment.
The workflow decides what happens.

Try a fictional software-support request. Jev classifies it, then a small set of code rules decides whether to prepare guidance, ask a question, or involve a person.

Simulation only. No real accounts, tickets, messages or transactions are connected. This illustrates my approach; it is not the production utility platform.

01 · The request

Start with a situation.

2–500 characters. Selecting an example only fills the box.

“Run live demo” sends this text to TypeSafe AI. Use fictional text only; leave out personal or confidential information.

02 · The rules

Keep the boundaries in code.

  1. Review consequential requests. Possible account changes, requests for a person, or attempts to bypass rules go to human review.
  2. Ask when the topic is unclear. Unsupported topics or category certainty below your threshold need clarification.
  3. Prepare guidance for the rest. Suggest an approved help article. No reply is sent by this demo.

Move this after a result to explore the tradeoff. It changes the local routing rule without another API call. Higher thresholds ask for clarification more often.

How to read these signals

Category certainty is Jev’s confidence estimate, not a guarantee that the category is correct. The other three signals estimate the chance that a condition is present.

A review signal of 20% or more triggers review here, including ambiguous cases. This is an illustrative policy, not a production-calibrated threshold. Changing the category slider cannot bypass it.

Read TypeSafe’s confidence documentation ↗