Jev 101
Start Guides
Search
中文 EN
Ops heat 1 · Quiet★ 0heat incomplete (no X)

Access Desk

Triages employee access requests: one Jev call answers typed questions, a code policy auto-approves only low-risk, well-justified, time-boxed requests and sends the rest to the system owner or manager with a recommendation.

Repo: samiulhuda360/access-desk ↗

What it is

An access-request desk that never quietly grants too much.

What it does: A request arrives by web form, Slack command or API. The desk adds directory context (role, team, manager, system sensitivity, current access), and one Jev call answers typed questions about fit, justification and risk. A code policy turns the probabilities into an action: auto-approve with an expiry date, or send to a person with Approve and Deny buttons. Allow and deny lists and sensitivity caps are enforced in code. Every decision goes to an audit log, and grants expire on their own.

Why it belongs in the showcase: A clear confidence-gated workflow with a three-way benchmark against a chat model and a static rules matrix.

Caveats: Invented data; simulated provisioning; the chat model was slightly more accurate.

Evidence

Public MIT repo; the author evaluates on 243 invented, labelled access requests and compares Jev, an OpenAI-compatible chat model and a role-to-system rules matrix on the 141 requests all three could assess: Jev auto-approved 0 risky requests (rules: 5), decision accuracy 73.8% (LLM 78.7%, rules 76.6%), p50/p95 latency 230/310 ms (LLM 1,106/15,858 ms), cost $0.0436 per 1,000 requests (LLM $0.1394); responses are cached so the evaluation replays offline.

Caveats

  • The evaluation set and the demo company are invented by the author
  • The default grant adapter is simulated and provisions nothing; real Okta, AWS IAM or Google Workspace integration is left to the user
  • The chat model baseline was slightly more accurate on this set; Jev automated fewer requests (20.6% versus 25.5%)

Unofficial case page. Stars / list position observed for heat; do not treat heat as a quality SLA. Review the upstream README before you depend on it.