Laya: The Open-Source Jev Alternative That's 7x Faster and Runs Under 1GB of RAM

Releases·Tool Release

laya-php: self-hosted PHP SDK for typed text classification with Laya, a Jev AI alternative

A new PHP SDK, laya-php, wraps the Laya text-classification engine, letting PHP apps route tickets, flag churn risk, and score urgency without per-token LLM API fees. Decisions return as typed PHP values (enums, ints, bools) with confidence scores, run self-hosted in 100+ languages. A Laravel-ready SDK lowers the barrier for PHP teams to replace LLM-based classification with a cheaper, private alternative.

A new PHP SDK, laya-php, wraps the Laya text-classification engine, letting PHP apps route tickets, flag churn risk, and score urgency without per-token LLM API fees. Decisions return as typed PHP values (enums, ints, bools) with confidence scores, run self-hosted in 100+ languages. A Laravel-ready SDK lowers the barrier for PHP teams to replace LLM-based classification with a cheaper, private alternative.

Releases·Tool Release

choosekit runs Jev-style local decisions via llama.cpp: 96.53% accuracy, 239 ms vs Jev's 368 ms

A developer unhappy with sending data to Jev's hosted API built choosekit, a small TypeScript package that extracts choices and probabilities from models running locally in llama.cpp. Tested with Qwen3.8 27B Q4 XL against Jev on SemIf's 144-task benchmark, both hit 96.53% accuracy, with the local setup faster (239 ms vs 368 ms median). The tool targets practitioners who want Jev-style decision extraction without third-party data exposure.

A developer unhappy with sending data to Jev's hosted API built choosekit, a small TypeScript package that extracts choices and probabilities from models running locally in llama.cpp. Tested with Qwen3.8 27B Q4 XL against Jev on SemIf's 144-task benchmark, both hit 96.53% accuracy, with the local setup faster (239 ms vs 368 ms median). The tool targets practitioners who want Jev-style decision extraction without third-party data exposure.

Products·concept

JevRev — a semantic decision layer that scores and filters LLM plans before more tokens are spent

JevRev pairs an LLM with 'Jev', a semantic layer that shortlists options, scores them in a loop, and monitors progress on ongoing work. Instead of being another coding agent, it acts as a decision layer beside the LLM, filtering plans and checking attention so cheap routing decisions aren't left to the model alone. It targets developers using LLMs for tasks that generate many candidate approaches, preventing wasted time and tokens on low-value directions.

JevRev pairs an LLM with 'Jev', a semantic layer that shortlists options, scores them in a loop, and monitors progress on ongoing work. Instead of being another coding agent, it acts as a decision layer beside the LLM, filtering plans and checking attention so cheap routing decisions aren't left to the model alone. It targets developers using LLMs for tasks that generate many candidate approaches, preventing wasted time and tokens on low-value directions.

Products·building

Weigh Swarm: research assistant with an "Inside the engine" view showing papers, claims, and model decisions

Weigh Swarm is a research assistant that answers questions from uploaded papers while letting users inspect how the answer was produced. LLMs write the cited synthesis, while a component called Laya makes bounded research judgments (claims, evidence selection). It targets researchers who want transparency into extraction, evidence, and model decisions rather than a black-box answer.

Weigh Swarm is a research assistant that answers questions from uploaded papers while letting users inspect how the answer was produced. LLMs write the cited synthesis, while a component called Laya makes bounded research judgments (claims, evidence selection). It targets researchers who want transparency into extraction, evidence, and model decisions rather than a black-box answer.

The Emerging Stack profiles TypeSafe's Jev, which bets there is a better approach than using highly capable language models for tiny decisions. The piece argues current AI usage over-applies conversational LLMs to tasks that may not need them.