Jev use cases directory
A public catalogue of real Jev use cases. Each row is a shipped job: classification, browser agents, guardrails, extraction, or retrieval, with the cost and latency the builder published.
| Use case | Pattern | Jev confidence | Builder | Cost | Latency | Result |
|---|---|---|---|---|---|---|
Classify 1,018 AI research papers Assign each paper one of 24 topics after a cheap summary step. | Classification | Hassan | $0.08 for classification, $3.99 for summaries | 256 ms median end to end per paper | 1,018 papers labelled into 24 topics | |
Zurich to London flights in 7.1 seconds Drive a live flight-search UI with Jev choosing the next browser action. | Browser agents | Gregor Zunic | $0.0039 | 1644 ms median decision, 7.1 s end to end | 5 actions to complete a Zurich to London search | |
Structured questions over images Ask a closed set of image questions and get typed yes/no and category answers. | Multimodal | Shin Megami Boson | Typed image answers instead of generated captions | |||
Label support tickets by queue Assign each incoming support ticket to one of billing, outage, or how-to after reading the subject and first message. | Classification | 100% | Community | |||
Route inbound support email Sort inbound customer emails into billing, product, and account-access queues. | Classification | 99% | Community | |||
State panel inbound billing queues Sort inbound customer emails into billing, product, and account-access queues. | Classification | 100% | Community | |||
Autoresearch feature discovery Propose Jev questions that turn tasting notes into numeric features for CatBoost. | Research | TypeSafe cookbook | Held-out RMSE 1.77 after five rounds, versus 3.09 for the mean baseline | |||
Double-checking citations Decide whether a quoted passage actually supports a claim. | Research | TypeSafe cookbook | Citations gated by support plus confidence | |||
Classifying RAG passages Score each retrieved passage before it reaches the answering model. | Classification | TypeSafe cookbook | Passage routing decided in code from Jev probabilities | |||
Date extraction from documents Ask Jev for the date parts named in a document, then resolve them in code. | Extraction | TypeSafe cookbook | Dates assembled in code from Jev-named parts | |||
Function calling from natural language Turn a trading request into a typed function name and closed-set arguments. | Function calling | TypeSafe cookbook | Typed function calls with published confidences of 0.91 and 0.94 | |||
Hierarchical classification Walk a deep taxonomy to the right leaf with parallel Choice questions. | Classification | TypeSafe cookbook | Leaf labels from parallel beam search over Choice probabilities | |||
Re-ranking legal passages Re-rank a 30-passage BM25 shortlist with one Jev question per pair. | Retrieval | TypeSafe cookbook | Top-1 5% to 18%, top-10 38% to 62% on 40 CLERC queries | |||
Guardrails for LLMs Screen every inbound and outbound message for hazards and harm. | Guardrails | TypeSafe cookbook | Hazard probabilities and a harm score per message | |||
13 questions in one Jev call Run a regulatory briefing as one batched Jev request instead of 13 singles. | Research | TypeSafe cookbook | 12.2x cheaper than 13 single-question calls | 10.0x faster than sequential singles | Same answers under both batching strategies | |
Pick one skill from 182 Choose at most one Hermes skill for an agent turn. | Function calling | TypeSafe cookbook | At most one skill name per turn from a 182-skill catalogue |