Article
CommunityJev Guides & Articles
LLMs Generate. Jev Decides. Inside a Java Invoice App
Mariano Barcia splits an invoice pipeline by semantics: the generative model extracts supplier, number, and amount, while Jev's two Choice questions handle supplier evidence and property selection, with policy left in Java.
- Format
- Article
- Reading time
- 2 min
- Published by
- Community
- Author
- Mariano Barcia
- Last verified
- Oct 2026
- Languages
- Java
Highlights
- Invoice Assistant at Mokapot Labs runs on The Pipeline Framework, an open-source framework for typed processing applications.
- Before: Gemma 12B extracted, classified supplier evidence, selected a property, and explained. After: it only extracts three fields.
- Jev answers two Choice questions — supplier evidence and property — and deterministic Java policy routes review versus vision analysis.
- TypeSafe published Python and JavaScript/TypeScript SDKs but no official Java SDK, so the team built a reusable Java client.
- The revised prompt shrank to extraction plus explicit do-nots for classifying, selecting, and explaining, which moved to Jev.
Watch out
A single production anecdote, not a benchmark: no accuracy, latency, or cost figures are given for Jev or the pipeline. The Java client is specific to The Pipeline Framework, and the article predates any official Java SDK, so check TypeSafe's current SDK list before copying the approach.
Questions about this entry
Short answers drawn from the linked project or page.
One Gemma 12B call did several jobs at once: extracting supplier, invoice number, and amount, classifying supplier evidence, selecting a property, and explaining the recommendation, with routing to human review or a vision model afterwards.
By semantic shape, not provider: extraction needs open-world generation, while supplier evidence and property selection have closed answer spaces. The old prompt literally told the LLM to behave as though the output space were closed, which described Choices.
Two Choice questions: supplier evidence (EXPLICIT_TEXT, STRONG_TEXTUAL_IDENTITY, or INSUFFICIENT) and the property selection from the supplied catalogue. The LLM is no longer asked to classify, select, explain a choice, or generate a note.
The open-world values: supplier name, invoice number, and total amount. When supplier evidence is insufficient, the existing vision model still performs the visual analysis after deterministic policy routes the case.
Not at the time of writing. TypeSafe publishes official Python and JavaScript/TypeScript SDKs, so the team instead built a thin, reusable Java HTTP client inside The Pipeline Framework.
Mariano Barcia at Mokapot Labs, using Invoice Assistant, a small application built with The Pipeline Framework, an open-source framework for typed processing applications that includes ordinary computation, AI inference, branching, a human Await boundary, and replay-safe Commands.
Uses AI is not a sufficient reason for two operations to belong to the same model call; their semantic shapes matter. The generative model generates, the decision model decides, and the application governs what happens next.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: How the agent stack moved in Jev's first three days
YouTube: Diogo Almeida tells swyx what System One is, why RLHF failed his automation goal, and why AI should eventually disappear.
X: The 10-step roadmap, summarised
X: Giving your agents a decision brain
More Jev resources like this
A Java and Python implementation of parallel constrained decoding: fill a whole JSON schema of booleans and enums in two forward passes, with calibrated confidence per field and output that is schema-valid by construction.
Jev Repos & SDKs#community#catalog-import#cobanov
Angel Nikolov's field report from three remote job boards: one Jev fan-out call per posting replaces keyword rules for sites, seniority, regions, work model, and benefits, with shadow mode before enforcement.
Jev Guides & ArticlesArticle#community#article#classification
Community
OpenRouter's Banking77 run: 3,080 support utterances, 77 intents, Jev 1.13 against Claude Opus 5 on the same one-line criteria, with latency, cost, and a confidence cascade.
Jev Guides & ArticlesArticle#community#article#evaluation
Community
The documentation site for OpenDecision, an open-source decision engine that runs locally and answers typed questions about text, JSON, documents and images, exposing Python, HTTP and TypeSafe-compatible interfaces with Choice, Noul, Score and Relation primitives.
Jev Guides & ArticlesDocs#community#docs#local
Community
MCP.so's explainer on Jev: how Choice, Score, and Noul differ from text generation, where a decision layer fits in retrieval, verification, routing, and ranking, and the four tools in the community Jev MCP server.
Jev Guides & ArticlesArticle#community#mcp#article
Community
DeepInfra's survey of the decision-model category: pointer-head architecture, RLCD calibration, latency and cost figures for Jev and its rivals, and a landscape table spanning TypeSafe, AWS, Convai, Kev, OpenAI, and Ollaya.
Jev Guides & ArticlesArticle#community#analysis#architecture
Community
From 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.