User repository on GitHub
TypeSafe AI - Jev - first look and experiment with its API + Google reviews experiments
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Practices & Patterns
Jev Reviews
An Astro 5 and TypeScript app that pulls Google Maps restaurant reviews from Apify into Supabase and classifies each one with Jev across five operational pillars, with generative models behind the same interface.
- Stars
- 2
- Language
- Astro
- Published by
- Community
- Author
- halfspin-qc
- Last verified
- Oct 2026
About this project
A high-performance, dark-themed web application built with Astro 5 and TypeScript to showcase bleeding-edge AI models, scrapers, and operational intelligence pipelines.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
Highlights
- Jev does the classification through Choice for sentiment, Score for the 1-5 pillars and Noul for recommends or has_complaints, so code reads probabilities.
- The classifier is swappable: OpenAI, Gemini and Groq sit behind the same engine, and the dashboard shows live latency and cost per inference.
- A deep-inspection drawer exposes question-level probabilities and confidences alongside extracted dishes and owner action items.
- Keys can be entered in the browser into sessionStorage instead of a .env file, which is gitignored.
Watch out
This is a demo workspace rather than a library: the README is a tour of the interface plus a setup guide, and it reports no accuracy measurement for the classification, only latency and cost. The figures, roughly 95-365 ms live and a fraction of a cent per review, are the author's own single-run numbers and move with review length and batch size. Setup is not one npm install: you need an Apify token and dataset, a Supabase project with a service-role key and an SQL migration, and the pricing figures quoted for Jev come from the README rather than from a billing page. The app reads the key from JEV_API_KEY rather than the TYPESAFE_API_KEY name used elsewhere in this ecosystem. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
Three of them, realistically: an Apify token and dataset id, a Supabase project with URL plus service-role and anon keys, and a model key. The README lets you skip .env entirely and enter keys in the browser through the API Keys button, which writes them to sessionStorage and nothing to disk. Without those services there is no data to classify, so npm run dev alone gets you the shell of the dashboard.
Unmeasured, as far as the README goes. It reports latency around 230 ms aggregated and a per-megabyte token cost, and it shows you the probabilities, but there is no accuracy number, no labelled benchmark and no comparison of Jev against the LLM backends on the same reviews. If classification quality is what you care about, the inspection drawer is where you would judge it by hand.
That is the variable name this project chose, alongside LLM_PROVIDER=jev and LLM_MODEL=jev-latest, while the OpenAI, Gemini and Groq paths all read LLM_API_KEY. Nothing else in the repo bridges the two names, so if you are following another Jev tutorial that exports TYPESAFE_API_KEY you will need to set both or map one to the other. The endpoint it posts to is the System One one, with jev-latest as the model alias.
The README describes five pillars scored one to five for food, service, cleanliness, atmosphere and value, a categorical sentiment Choice, and Noul questions such as recommends, has_complaints and has_praise, which is how it produces typed probabilities instead of prose. The stated advantage is single-pass evaluation at a fraction of the cost, with the Gemini, Groq and OpenAI backends available when you want a different trade-off for the same task.
Apify pulls Google Maps reviews, a client library unrolls them from both flat datasets and nested place records, and Supabase PostgreSQL stores them with JSONB classification columns, indexes, RLS policies, and summary stats. Classification runs through one engine in src/lib/llm.ts that Jev, OpenAI, Gemini, and Groq all sit behind, so switching provider changes the trade-off without changing the interface.
The README quotes live responses at roughly 95 ms to 365 ms and a typical classification at about $0.00002 to $0.00004, on pricing of $0.042 per million input tokens with output free. The dashboard aggregates that to roughly 230 ms average. These are the author's single-run figures and move with review length and batch size.
npm install, copy .env.example to .env and fill in the Apify token and dataset id plus the Supabase URL, service-role key, and anon key, then npm run dev and open the jev-reviews route on localhost:4321. A setup guide covers the Apify scraper run and the Supabase SQL migration. The repository is MIT and .env is gitignored.
More Jev resources like this
38GitHub stars
Local unofficial /v1/systemone adapter that serves Choice, Score, and Noul probabilities via first-token LLM judgments or a fine-tuned ModernBERT scorer, including Japanese docs.
Jev Practices & Patterns#community#awesomejev#catalog-import
Community
Argos1111/jev_localPython
Argos1111/jev_localPython24GitHub stars
Unofficial laptop study of parallel constrained decoding: prefill once, share the KV cache across fields, and decide all schema fields in one batched forward pass without writing JSON.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
rorshopping/jev-on-a-laptopPython
rorshopping/jev-on-a-laptopPython23GitHub stars
Live race rig that labels 1,000 real app reviews with Jev typed questions against Gemini JSON output on identical batches, then replays recordings and re-ranks rows locally without extra model calls.
Jev Practices & Patterns#community#jevbest#catalog-import
22GitHub stars
Offline evaluator that checks whether classifier confidences match reality and converts mistake costs into a deployable threshold, emitting an HTML report plus machine-readable threshold files.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
rlaope/jevalPython
rlaope/jevalPython16GitHub stars
Vendor-neutral harness that tests whether Jev reranking improves a small RAG stack, using frozen hybrid BM25 plus BGE-M3 candidates and isolating retrieval, rerank, and generation stages.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
erendikmenn/jev-rag-benchmarkPython
erendikmenn/jev-rag-benchmarkPython37GitHub stars
Minimal Python recipe that turns any causal LLM into a Jev-style classifier by prefilling choice_index and constraining sampling to legal indexes in one forward pass.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
sgoedecke/system-onePython
sgoedecke/system-onePythonFrom the Jev guides
Original write-ups on what this entry is for.
Most Jev accuracy problems are state problems. What belongs in a state, what is a pre-made judgment to leave out, and why naming the slice you ask about matters more than the wording.
Jev Practices & Patterns#state#design#primitives
Read article
A Noul returns one number and no separate confidence. How to read the value, pin its boundary with criteria, split compound questions, and set a per-action threshold that matches the cost of being wrong.
Jev Practices & Patterns#noul#primitives#confidence
Read article
A Score is a weighted position, not a ruler. Why the number can land between levels, why levels are judged separately, and how to combine several atomic scores in code instead of one crowded rubric.
Jev Practices & Patterns#score#primitives#scoring
Read article
From the Jev community
Posts from builders shipping with Jev right now.