Use-case gallery

Remote browser agent loop (Stagehand + Jev)

Observe a page, send the a11y tree as state with actions as questions; Jev picks the next operation/targets, Stagehand executes—~$0.001 per task at near-instant speed.

choice · noul · score

Block-speed buy/sell trading bot (Monad + Kuru)

Jev chooses buy or sell from an asset price feed each ~300ms Monad block and places post-only limits on Kuru’s on-chain book.

choice

Real-time Doom bot (~10 QPS)

TypeSafe launch demo: Jev drives Doom at ~10 calls/sec (~$7/hour) for real-time game decisions from structured game state.

choice · noul · score

Wikipedia link race (high-cardinality Choice)

Wikiracing demo: among hundreds/thousands of outbound links, Choice picks the next page without hallucinating invented links.

choice

Open-source Jev-ify any HF model (SimpleJev)

Builder ships SimpleJev.ai: open-source library to turn any Hugging Face model into Jev-like Choice/Noul/Score API, including vision.

choice · noul · score

Semantic search replace/supplement embeddings

Replace or supplement embeddings in RAG with semantic search, scoring, and ranking over candidates.

score · noul · choice

Query–candidate relevance scoring

Score how well each retrieved candidate answers the query before generation.

noul · score

Pairwise result reranking

Rerank shortlists with pairwise comparisons using TypeSafe questions.

choice · noul

Cross-encode query and candidate

Cross-encode query+candidate for higher precision than bi-encoders alone.

noul · score

Select useful context for downstream AI

Pick which snippets are useful enough to pass into an LLM or agent step.

noul · choice

Systematic review paper screening

Screen papers against inclusion/exclusion criteria for systematic reviews.

choice · noul

Qualitative theme labeling

Label interview transcripts, surveys, and field notes with predefined themes.

choice

Claim–citation support check (science)

Check whether cited passages support claims in manuscripts and summaries.

choice · noul

Flag missing methodological details

Flag missing controls, dataset descriptions, and experimental settings.

noul

Research knowledge-graph entity linking

Identify entities and relationships across papers linking findings to passages.

choice · noul

Custom LLM model router

Build a router that chooses which LLM receives each prompt using Jev.

choice · score

Intent and domain classification for routing

Classify intent and domain to set routing rules and thresholds.

choice

Estimate difficulty and risk of prompts

Estimate prompt difficulty/risk before escalating to expensive models.

score

Escalate hard requests to expensive models

Escalate only requests that need a more capable/expensive model.

noul · choice

Jailbreak and prompt-injection detection

Detect jailbreaks and prompt injection on LLM inputs at a fraction of LLM cost.

noul · score

Policy violation and sensitive-data checks

Identify policy violations and sensitive-data exposure on I/O.

noul

Tool-call error detection

Detect tool-call errors and response-quality failures in real time.

noul · score

Structured guardrail logging

Log structured check results and probabilities for harness failure tracing.

noul · score

Semantic code linting in CI

Add semantic lints for team coding conventions and writing guidelines in CI.

noul · choice

Probabilistic NL features for ML

Extract probabilistic features from natural language for predictive models.

noul · score

Autoresearch feature discovery loop

Propose feature definitions and evaluate predictive value on held-out truth.

noul · score

Resume and interview criteria scoring

Evaluate resumes/applications/interview feedback against job criteria.

score · noul

Candidate–role matching and routing

Match candidates to roles and route to hiring managers; escalate uncertain cases.

choice · noul

ICP lead matching

Match company profiles and inbound messages to an ideal customer profile.

noul · score

Buyer intent and pain-point detection

Detect buyer relevance, pain points, and purchase intent; prioritize leads.

noul · score

Support ticket multi-axis classify

Classify tickets by issue, product area, and customer intent.

choice

Call transcript issue/commitment extraction

Extract issues, commitments, and follow-ups from call transcripts.

choice · noul

Urgency, frustration, churn, refund signals

Detect urgency, frustration, churn risk, and refund requests on tickets.

noul · score

Support queue and automation routing

Route cases to team, queue, or automated workflow.

choice

Verify support replies vs policy

Verify support responses against policies and the customer request.

noul · choice

Insurance FNOL classification

Classify first-notice-of-loss, adjuster notes, and supporting docs.

choice

Claim complexity, missing info, fraud flags

Detect claim complexity, missing information, and potential fraud indicators.

noul · score

Straight-through vs specialist claim routing

Prioritize claims for STP or specialist review; escalate uncertain cases.

choice · noul

AML/KYC suspicious narrative scoring

Evaluate transaction narratives, KYC docs, and alert histories for suspicion.

noul · score

Entity match across inconsistent names

Match entities across inconsistent names, profiles, and records.

score · choice

Financial alert prioritization

Prioritize alerts by risk, relevance, and evidence quality.

score

Contract and filing classification

Classify contracts, policies, regulatory filings, and marketing claims.

choice

Missing clauses and prohibited claims

Detect missing clauses, prohibited claims, and policy violations.

noul

Legal requirement verification

Verify documents against explicit legal or compliance requirements.

noul · choice

Product listing normalize/classify

Classify and normalize product listings across inconsistent seller catalogs.

choice

Product attribute extraction

Extract product attributes from titles and descriptions.

choice · noul

Counterfeit, review abuse, prohibited listings

Detect prohibited listings, counterfeit signals, review abuse, policy violations.

noul · score

Nuanced community moderation

Apply company-specific criteria to decide which posts meet moderation standards.

choice · score

Toxicity, spam, PII, opt-out detection

Detect toxicity, harassment, spam, fraud, unsafe advice, PII, opt-outs.

noul · score

Severity×confidence allow/warn/block

Combine severity and confidence to allow, warn, review, or block.

score · noul

Ad brand safety and suitability

Classify brand safety and audience suitability for creatives and placements.

choice · score

Ad regulatory claim compliance

Check regulatory compliance and prohibited claims in ads.

noul

Creative quality and LP alignment

Evaluate creative quality and ad-to-landing-page alignment.

score · noul

Game chat abuse moderation

Moderate in-game chat; detect abuse, toxicity, suspicious behavior.

noul · score

Player frustration and engagement scoring

Annotate content and score frustration or engagement from reports/reviews.

score

Player churn and support routing

Detect churn signals and route player-support requests.

noul · choice

Incident-to-risk probabilistic indicators

Convert incident reports and vendor assessments into probabilistic risk indicators.

noul · score

Underwriting risk type classification

Classify risk types and detect suspicious characteristics for underwriting.

choice · noul

Demand forecasting semantic enrichment

Enrich forecasts with purchase intent, urgency, and product interest from text.

noul · score

Supply/competitive pressure themes

Detect supply concerns, competitive pressure, and emerging demand themes.

noul

KG relationship and entity typing

Classify relationships and entity types when annotating knowledge graphs.

choice

Contradiction detection in records

Detect contradictions between records or claims in a knowledge graph.

noul

Self-consistency Noul review gating

Route uncertain Noul probabilities to human review while keeping underlying values visible.

noul

Self-consistency Choice moderation

Add uncertain outcome to moderation; compare label agreement with share of automatic actions.

choice

Parallel regulatory briefing questions

Batch many questions (e.g. 13 GDPR briefing Qs) in one call—cheaper and faster with same answers.

choice · noul · score

Legal BM25 + TypeSafe rerank (CLERC)

Rerank BM25 shortlists with one question per query–candidate pair to lift legal retrieval accuracy.

noul

Line-by-line semantic document search

Score ToS line IDs against a plain-language query; Noul checks whether an answer exists.

choice · noul

Recover Markdown structure from plain text

Stitch hard-wrapped lines and classify blocks (heading, list, code, callout) with companion questions.

choice · noul

Map NL trading requests to typed functions

Map function names and closed-set args to confidence-aware questions for ordinary typed functions.

choice · noul

Agent skill suggestion from large catalog

Rank Hermes catalog skills and decide whether a turn needs a skill at all.

choice · noul · score

Beer catalogue entity alignment

Score candidate pairs as merge / leave unlinked / hand to curator across two catalogues.

score

Classify RAG passages for injection/contradiction

Score retrieved passages; keep/flag contradictions; drop hidden instructions or injections.

noul · choice

Double-check citations against source

Choice whether quote context supports the claim; confidence flags for human review.

choice

LLM I/O hazard guardrails battery

Noul hazards + severity Score on every LLM input/output; threshold to pass/review/block/support.

noul · score

SDE cascade mini→verify→reason

Two-stage structured-data-extraction cascade for big-model quality at lower cost.

noul · choice

Date part extraction then code resolve

Ask TypeSafe for named date parts, resolve/validate in code with confidence-based review.

choice · noul

Regex candidates + TypeSafe span pick

Regex finds emails/phones/amounts; TypeSafe selects the requested span for verbatim normalize.

choice

Hierarchical beam-search classification

Classify through deep patent/retail/biomed/code hierarchies via beam search on Choice probs.

choice

Autoresearch → CatBoost features

Autoresearch loop proposes TypeSafe questions as numeric features for a CatBoost regressor.

noul · score

SEC report industry Classification + confidence

Classify SEC annual reports into industry groups; use confidence to fall back to broader division.

choice

Speculative fan-out many questions

Send many speculative questions in one call; code decides which answers matter.

choice · noul · score

Confidence-gated act vs escalate

Use answer for what and confidence for whether to act or escalate.

choice · noul · score

Composite weighted scoring

Break complex judgment into atomic scores; combine with code-owned weights.

score

Intent routing to code/LLM/human

Classify requests and route to deterministic logic, specialist LLM, or human.

choice

Smart-home request evaluation demo

Demo: evaluate smart-home user requests with typed System One questions.

choice · noul

OCR screen → next computer action

Community typesafe-computer-use: OCR the screen then classify the next desktop/computer action (~$0.0002/step).

choice · noul

Emulator Mario from structured state

Community typesafe-mario: play Super Mario Bros from structured emulator state (same shape as Doom demo).

choice · noul

Staged code-review workflow (jev-review)

Community jev-review: staged code-review workflow with local dashboard using typed review questions.

noul · score · choice

Software factory orchestration (foreman)

Community foreman: orchestration layer for a software factory making typed progress/quality decisions.

choice · noul

LangChain model-routing middleware

LangChain TypeSafe middleware: Jev assesses the user message and selects fast vs powerful model for the run.

choice

LangChain AutoMode dangerous-tool guard

AutoModeMiddleware uses Jev to classify risky tool calls (e.g. bash) and block before execution.

noul · choice

Email triage at scale (Ryan Vogel)

Public LangChain writeup cites Ryan Vogel doing email triage at scale with Jev decisions.

choice · noul · score

Support department routing (Cloudflare Workers AI)

Cloudflare docs example: route support requests to billing/tech/other with Choice on login/payout failures.

choice

Structured refund review (Cloudflare)

Evaluate refund requests against order+policy with Noul/Choice/Score in Workers AI example.

noul · choice · score

Account risk escalation (Cloudflare)

Score account risk and Noul whether activity warrants immediate escalation.

score · noul

Newsletter sponsor inquiry filter

Flavio Copes pattern: Noul for real sponsorship inquiry, Choice for category, Score for ask specificity.

noul · choice · score