Astrolabe
Local tool where Jev reads every accepted paper's abstract at NeurIPS, ICLR or ICML and judges relevance to your research topic.
Overview
Astrolabe helps researchers find the papers that matter to them among thousands accepted at a conference. Instead of keyword matching or embedding similarity, Jev reads each paper's abstract and returns the probability that it falls under your described topic. The README reports judging all 5,857 NeurIPS 2025 papers in about 2 seconds for about $0.08.
Papers are shown as constellations grouped by the conference's subject areas, with matches lit up. Because answers are probabilities, a threshold slider changes strictness without calling Jev again, and results export to CSV. It covers NeurIPS, ICLR and ICML for 2024 to 2026; NeurIPS 2026 is judged on titles only until abstracts are public, which the README says lowers F1 from 0.88 to 0.60. It runs locally with your own OpenRouter or TypeSafe key, stored in the browser and passed only through your local server. Docs are in Korean and English.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Does this paper fall under the user's research topic? | yes · no |
How it works
- Download conference paper data
- Enter your Jev key and a detailed topic
- Jev judges every abstract
- Adjust the threshold slider and export CSV
Strengths
- Whole conference in seconds for cents, per the README
- Threshold changes need no re-calls
- Runs locally with your own key
Caveats
- Title-only judging is much less accurate
- Key stored in plain text in localStorage
- Not meant for public deployment
Getting started
git clone https://github.com/yongukpark/astrolabe.git && cd astrolabe && npm install && npm run data && npm run devAt a glance
- Type
- Project
- Status
- beta
- Language
- TypeScript
- Runtime
- local Node web app
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Astrolabe is a local research discovery tool that uses Jev to read and classify thousands of conference papers by how relevant they are to a user-defined research topic.
hellogumbo/awesome-jev
Frequently asked questions
What does Jev decide?
Whether each paper is relevant to your topic, as a probability.
How much does a search cost?
About $0.02 to $0.09 per conference, per the README.
Which conferences?
NeurIPS, ICLR and ICML for 2024 to 2026.
Related projects
Assembles source-linked evidence for an AI task within a token budget using Jev relevance judgments and diversity-aware selection.
ProjectbetaPythonResearch agent where GPT plans and Jev finds evidence in parallel across document regions with Choice and Noul.
Browser extensionresearchTypeScriptDeepSeek Harness paper-reading workbench whose PDF search reranks keyword candidates with Jev, falling back to keyword order.
BenchmarkbetaTypeScriptChinese-language Zotero AI reading assistant that uses Jev judgments to classify papers in your library.
LibraryproductionJavaScriptMulti-user Telegram bot that judges every new post in public channels with Jev templates and routes matches to your DM or channels.
ProjectbetaPythonTelegram channel X-ray: Jev judges up to 1,500 public posts for kind, paid ads, clickbait and pressure, producing a shareable card.
ProjectbetaJavaScript