Things people built with Jev.
An unofficial catalogue of projects, recipes and code on TypeSafe’s System One models.
- pg-jevA PostgreSQL extension that lets you write a WHERE clause in plain language: jev(people, the name is European) is an ordinary boolean function, so it joins, groups and sorts like any other. Each row is judged by Jev, batched many rows per request and cached per row, with jev_prob, jev_choice and jev_score for ranking and classifying. No embeddings, no vector column.
- Jev DetectorA free AI-slop detector. Paste up to 10,000 words and Jev checks every sentence against known AI-writing patterns, then highlights the ones worth a second look with the pattern it matched, a strength score and a confidence. Framed as a writing aid rather than a verdict, with the Wikipedia rubric it draws on linked. No login.
- Jev Moderation BotA Discord moderation bot in Python that asks Jev whether each message is spam or a scam link, deletes offenders on the spot and escalates from a warning to a timeout. A /profile command scores a member across scam risk, spamminess, toxicity and helpfulness, and pardons are kept as precedents so the same false alarm is not repeated.
- skillrankerA standalone Rust CLI that puts TypeSafe.ai's Jev at the center of skill selection: Jev evaluates your agent's live context, compares the available skills, and estimates which ones fit the next step. SkillRanker supplies the session integration, local safeguards, and inspectable feedback around it.
- Jev UltrafastA browser agent from Browser Use where Jev decides every step. From a numbered table of the page's elements it picks the operation and the target in a single request, and a small LLM is called only when text has to be typed. Its demo runs a Google Flights search from one plain-English goal in about seven seconds.
- Autoresearch feature discoveryA loop that proposes new questions, turns free text into numeric features with them, and uses a regressor's mistakes to decide what to ask next.
- Classification using confidenceSort company annual reports into fine-grained industry groups with one Choice each, and use the answer's own confidence to fall back to the broader division when the model is unsure.
- Classifying RAG passagesJudge each retrieved passage before it reaches the answering model: keep what helps, flag what contradicts the question, and drop anything carrying a hidden instruction.
- Date extractionPull absolute and relative dates out of documents by asking for the parts that are named, then resolving and validating them in code, with low-confidence cases sent for review.
- Double-checking citationsCheck whether a quoted source really backs the claim it is attached to. One question gives the verdict, and its confidence decides which citations a person should look at.
- Function callingTurn a plain-English request into a call to an ordinary typed function. Each function name and each closed-set argument becomes a question, so the answer is always something your code can dispatch.
- Guardrails for LLMsScreen what goes into and comes out of an LLM app in a single request: describe the hazards, score how bad complying would be, and let your code decide to pass, review, block or reroute.
- Hierarchical classificationClassify documents into deep category trees, such as patents, retail products, biomedical topics and source code, by running a beam search over the probabilities of each level.
- Knowledge graph entity alignmentDecide which candidate pairs from two product catalogues are the same thing. A single Score question with three levels maps straight onto the three actions: merge, leave apart, or send to a curator.
- Line-by-line searchSemantic search over a long terms-of-service document in one request: a Choice picks the line that answers a plain-language query, and a Noul checks whether the document answers it at all.
- Parallel questionsAsk a whole briefing's worth of questions about one long document in a single call instead of one call each. Same answers, at a fraction of the cost and time.
- Pre-parsed value extractionLet a regex find every candidate email, phone number or amount, then have the model select the right one. The output is always a verbatim value from the text: select, do not generate.
- Re-rankingTake a keyword-search shortlist and ask one question per query and candidate pair to reorder it. Worked through on legal case retrieval, where it lifts the right passage towards the top.
- SDE cascadeA two-stage cascade for structured data extraction: a small model extracts, a verification step checks, and only the doubtful cases go to a large reasoning model. Most of the quality for a fraction of the cost.
- Self-consistency: choicesAdd an explicit uncertain outcome to moderation decisions, then weigh how often labels agree against how much gets actioned automatically.
- Self-consistency: noulsSend borderline probabilities to a human reviewer instead of forcing a yes or no, while keeping the underlying values in view so the routing stays explainable.
- Skill suggestionPick at most one skill for an agent's turn from a catalogue of well over a hundred. One request ranks them all and asks whether any is needed; a second reads the top few closely and may reject them all.
- Structure recoveryRebuild Markdown from plain text that lost its formatting. One request rejoins hard-wrapped lines; a second labels each block as heading, list, code or callout.
End of catalogue. Add yours →