Examples
38 use cases, grouped the way TypeSafe’s use-case map ↗ groups them. Each has a real recorded answer and opens in the playground.
AI automation software
Code owns the workflow; Jev makes the semantic calls, a million times over.
Support ticket
Route a ticket to the right team, and know when a person should check first.
Email triage
Sort an inbox: what kind of email, how urgent, how big the opportunity.
Lead scoring
Score a signup against your ideal customer and spot who can sign.
Résumé match
Check a candidate against the must-haves before anyone spends an hour.
Meeting notes
Tell a meeting that decided something from one that just talked.
App review
Turn a store review into a routed, prioritised signal.
Invoice extraction
Pull the fields your accounts system needs out of an invoice. Code finds the candidate amounts; Jev picks which is which.
Comment replies
Find the comments worth a creator's reply among hundreds.
Insurance claim
Triage a first notice of loss: settle it automatically, or send an adjuster.
Contract review
Check a clause against an explicit requirement and escalate what fails.
Real-time applications
Decisions in about 150 ms — fast enough for games, UIs and control loops.
Fraud check
Turn a transaction's signals into a risk your payment rules can act on.
Product match
Score a product against a live shopping session; your recommender does the ranking.
Market news
Read a headline into a signal. Jev feeds the risk engine; it never places the trade.
Game NPC
Decide an NPC's next move from game state, fast enough to run every frame.
Robot action
Pick a robot's next action from sensor state; safety interlocks still sit in code.
Map-reduce over big data
Cheap enough to classify, rank and extract features across huge datasets.
Viral post
Read a draft post the way an editor would before it goes out.
Feedback tagging
Tag thousands of survey answers with category and churn risk, for pennies.
Search ranking
Score a search result against what the searcher actually wants — a score, not a summary.
Paper screening
Screen abstracts against a systematic review's inclusion criteria.
AML alert
Prioritise money-laundering alerts so investigators see the real ones first.
Demand signal
Turn sales notes into features a forecasting model can use.
Knowledge graph
Type the relationship between two entities before it becomes an edge.
Help-centre search
Find the article that answers a question in plain words, not the one that shares its keywords.
Universal verification
Check prompts, answers, tool calls and documents for specific failure modes.
Comment moderation
Decide what a comment is, whether it threatens anyone, and how bad it is.
Prompt guardrail
Check a user prompt before your assistant ever sees it.
Tool-call gate
Let an agent act, but stop the calls that go beyond what it was asked to do.
Code review risk
Flag the pull requests a senior engineer must look at.
Phishing check
Catch the credential-stealing email before anyone clicks.
LLM answer check
Judge another model's answer before a user sees it. A chatbot produces; Jev judges.
Semantic lint
Check code against a team convention no regex can express. Runs in CI.
Listing check
Spot counterfeit and policy-breaking listings across a marketplace.
Ad review
Check ad copy for prohibited claims and brand safety before it serves.
Harness engineering
Make an agent harness smarter: routing, context, next steps and tool choice.
Model routing
Send each request to the cheapest model tier that can do it well.
RAG passage check
Keep only the retrieved passages that actually answer the question.
Context compaction
Decide what an agent keeps in its context window as a conversation grows.
Agent routing
Send each request to the right specialist agent: classify, route, execute.
Agent next step
Make the many small "what now?" decisions inside an agent loop, fast.
Tool selection
Pick the right tool for an agent's next call, and spot calls with side effects.
Decision shapes
The ten shapes TypeSafe names, and which use cases here show each one.
| Shape | Reach for it when | Try it |
|---|---|---|
| Classification | One known category should winIntent, topic, department, risk type, entity type | Email triageApp reviewComment repliesPaper screeningKnowledge graph |
| Detection | You need a probability that one property is presentSpam, fraud, urgency, jailbreaks, sensitive data | Comment moderationPrompt guardrailMeeting notesPhishing checkFraud checkMarket newsListing check |
| Scoring | The answer belongs on an ordered rubricSeverity, relevance, quality, frustration, suitability | Lead scoringCode review riskRésumé matchViral postAML alert |
| Routing | A category selects the next code pathTool use, escalation, model routing, support queues | Support ticketModel routingAgent routingAgent next stepTool selectionGame NPCRobot actionInsurance claim |
| Search | You need to find items that match a natural-language querySemantic search, document discovery, candidate generation | Help-centre search |
| Retrieval | A workflow needs the most relevant context or recordsRAG context, evidence retrieval, knowledge lookup | RAG passage checkContext compaction |
| Ranking | Items need to be ordered by semantic relevance or qualitySearch results, recommendations, candidate prioritisation | Search rankingProduct match |
| Verification | An artifact must be checked for specific failure modesCitation support, policy violations, tool-call errors, response quality | Tool-call gateLLM answer checkSemantic lintContract reviewAd review |
| ML feature extraction | A downstream classical ML model needs semantic signalsPurchase intent, product interest, competitive pressure, churn signals | Feedback taggingDemand signal |
| Structured data extraction | Known fields must be recovered from unstructured inputCandidate attributes, order fields, document labels | Invoice extraction |
From the TypeSafe docs
The requests the documentation uses, with the numbers it prints. The lessons and the Limits page are built on these.
Support & routing
First run
The whole contract in one click: three question types, one request.
2 variantsSupport triage
Five Choice questions in one call, two of them speculative.
4 variantsVoice banking intents
One intent question, two different bars to act on it.
4 variantsIntent routing
Send each customer message to the right handler — a lookup, a specialist LLM or a person — using thresholds in your code.
3 variantsSmart home assistant
Ask every question up front, including the ones that will not apply, and filter in code.
4 variantsGuardrails & verification
LLM guardrails
Four hazard Nouls and a severity Score, on both sides of an LLM call.
7 variantsCitation check
Does the cited section actually support the claim, or just contain the quote?
4 variantsRAG passage gate
Judgment beats similarity: four Nouls catch a planted injection and a passage that refutes the query, which cosine ranking missed.
3 variantsScoring
Limits — things Jev gets wrong on purpose
Counting
Jev recognises the shape of an answer rather than tallying. Count in code.
2 variantsStructural invariants
Two phrasings of one question need not agree, and probabilities across questions need not sum to 1.
2 variantsNumeric score levels
A level is judged on its description alone. Number the levels and there is nothing to judge.
2 variantsInstructions hidden in the state
State is data, and Jev does not treat it as hostile. Say in the criteria what the state is.
2 variantsDate arithmetic
Jev reads dates as text, not as quantities. Extract the parts; do the arithmetic in code.
2 variantsEvery replay says where its numbers came from: Jev Lab’s own run of the exact request, with the model id and date, or numbers quoted from docs.typesafe.ai, with the page. They show the shape of a result, not a benchmark, and a live run can differ.