Article
CommunityJev Integrations & Adapters
Spam and Auto-Reply Detection in There There
Freek Van der Herten describes replacing a fifteen-language subject-prefix list with Jev in There There's helpdesk: three questions in one call, header checks first, raw probabilities stored next to each thresholded verdict.
- Format
- Article
- Reading time
- 3 min
- Published by
- Community
- Author
- Freek Van der Herten
- Last verified
- Oct 2026
- Languages
- PHP
- Capabilities
- Requires API key
Highlights
- Jev support landed in the 1.x branch of the Laravel AI SDK a day before the post, and There There started using it for spam detection the same day.
- Mail headers are checked first, and the subject-prefix lists Jev replaced covered fifteen languages plus nine bounce patterns.
- An InboundJudgement enum carries each question with its threshold: IsAutoResponse 0.75, and IsBounce and IsSpam 0.9; true and false descriptions are optional.
- All unsettled questions go out in one request with a 10-second timeout; the body is truncated with Str::limit(..., 10_000) and failures are logged and skipped.
- Measured: three questions on one mail take 639 ms, about 48 per second in parallel, at $0.042 per million input tokens - roughly 36 cents a month.
Watch out
One team's production notes rather than a benchmark: the 639 ms and 48-per-second figures come from There There's own setup with no tuning, and the calibrated-probability claim is TypeSafe's. The $0.042 per million input tokens price and the cents-per-month estimate are specific to that workload.
Questions about this entry
Short answers drawn from the linked project or page.
Helpdesk mail that is not from customers: out-of-office replies, bounces, subscription confirmations, DMARC reports, and spam. Header checks and a growing subject-prefix list handled some; everything they could not answer now goes to Jev before an LLM writes titles and summaries.
Classification::of(state)->question(...)->classify() returns answers by name: a Boolean gives a probability, so urgent came back 0.94 with isTrue(threshold: 0.8); a Choice gives the pick, probabilityOf per option, and confidence; a Score rates on levels and can land between them.
As a string or an array whose parts have names. The refund example combines subject, message, an order with its charges, and a refund policy into one state, still a single thing being judged.
Each InboundJudgement case carries its own threshold and is applied with isTrue(threshold). Raw probabilities are stored next to the verdicts so a threshold can change later and you can see what it would have done; workflow conditions then read a column without calling Jev.
Configure Jev like any other provider in config/ai.php with a TYPESAFE_API_KEY in your env file, then use the Classification facade with Boolean, Choice, and Score questions, a 10-second timeout, and the message body truncated before sending.
Classifying one mail with three questions took 639 ms, and the pipeline processes around 48 per second in parallel. At $0.042 per million input tokens with free output tokens, the post estimates four hundredths of a cent per mail, about 36 cents a month.
It reports one helpdesk's operational numbers, not an accuracy benchmark; the claim that a 0.9 probability is right about nine times in ten is TypeSafe's description of calibration. Thresholds here are the author's choices, and the failure path just logs and moves on.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: 1,565 emails, six decision models
X: Fifty emails in under two seconds
More Jev resources like this
The Laravel AI SDK's docs carry a Classification section covering classifying images, yes or no decisions, and choosing from collections, and list TYPESAFE_API_KEY among the provider credentials in config/ai.php.
Jev Integrations & AdaptersDocs#community#php#laravel
Community
A follow-up on There There's Jev use: workflow conditions described in plain language become a Boolean question in the same inbound classification request, and Jev picks a ticket's assignee from a curated list.
Jev Integrations & AdaptersArticle#community#php#laravel
Community
Filament plugin that filters admin table rows by a plain-English statement: each row gets a Jev Noul probability that the statement is true, and rows at or above the threshold stay. Scores are cached by content, so reruns of the same statement are instant and free.
Jev Integrations & Adapters#community#laravel#filament
Community
vibefilter/filamentPHP
vibefilter/filamentPHPA PHP decision engine for Laravel that replaces open-ended prompts with typed primitives: Noul for binary likelihoods, Choice for categorical classification and Score for bounded ratings, each returning calibrated probabilities.
Jev Integrations & Adapters#community#github#php
Community
obelaw/systemonePHP
obelaw/systemonePHPOpen-source AI email assistant (Next.js, Postgres, 12k+ stars) with a decision-model layer that can run on openrouter:typesafe/jev-1.13: Jev-backed questions cover rule selection, cold-email detection, thread status, recurring patterns, unsubscribe pages, and sender categorization, each with evals.
Jev Integrations & Adapters#community#email#gmail
Community
elie222/inbox-zeroTypeScript
elie222/inbox-zeroTypeScriptA Laravel and Livewire workshop demo that runs a Twitch-style live chat through four moderation setups in parallel: a conventional LLM agent, the Decisions API decision model, a local Laya model and OpenAI Moderation, comparing accuracy, latency and cost.
Jev Demos & Experiments#community#github#php
Community
beerandcodeteam/workshop-jevHTML
beerandcodeteam/workshop-jevHTMLFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.