JevPulse
YouTube comment analyser: an LLM writes a per-video rubric once, then Jev scores every comment and code computes consensus.
Overview
JevPulse analyses what an audience is saying in a YouTube video's comments. It replaces conversational summarisation with a two-stage design. First, Gemini Flash is called once per video with the title, description, transcript excerpt and sample comments to produce a rubric of closed questions, each with affirmative and negative boundary criteria to prevent drift.
Then every comment is evaluated with Jev in batches sharing one state. Universal questions run on all videos: whether the comment is an opinion, its type (praise, criticism, question, suggestion, agreement, disagreement, correction, experience or other), a 3-point specificity score and whether it makes a verifiable claim. Video-specific yes/no questions come from the rubric. Application code then aggregates, thresholds and computes consensus deterministically. A live demo is hosted on Vercel. It suits creators and analysts who want per-comment evidence instead of a single LLM summary.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Is this comment a substantive opinion (not a greeting, spam or timestamp)? | — |
| Choice | What type of comment is this? | praise · criticism · question · suggestion · agreement · disagreement · correction · experience · other |
| Score | How specific is the comment's argument? | — |
| Yes / no | Does the comment affirm this video-specific rubric criterion? | — |
How it works
- Fetch video metadata, transcript excerpt and comments
- LLM synthesises a video-specific rubric once
- Batches of comments are sent to Jev with universal and rubric questions
- Code aggregates probabilities into consensus metrics
Strengths
- Every comment is evaluated individually
- One LLM call per video; bulk work done by Jev
- Deterministic aggregation in code
Caveats
- Depends on both Gemini and Jev APIs
- Rubric quality depends on the LLM's first pass
- Demo project
At a glance
- Type
- Project
- Status
- prototype
- Language
- TypeScript
- Runtime
- web app
- Has code
- yes
Use cases and integrations
How curated lists describe it
Analyze the YouTube comments to gain insights into what the audience is saying from any video link.
logicrw/awesome-jev-projects
Frequently asked questions
What does Jev decide?
For each comment: whether it is an opinion, its type, its specificity, whether it makes a claim and how it stands on each rubric criterion.
What is the LLM used for?
Only to design the per-video rubric once; it does not read every comment.
How is consensus computed?
Deterministically in application code from Jev's probabilities.
Related projects
Self-hostable Reddit buyer-intent finder that has Jev judge every post and comment it reads, with a written reason and cost per lead.
ProjectproductionTypeScript- jev-seo★ 610
Live SEO audit from one homepage URL: crawl, 52 rules, Core Web Vitals and Jev page judgments, reported as PDF, XLSX and Markdown.
PluginbetaPython Scores HeyReach LinkedIn leads and first-message variants with Jev, picks a message per lead and writes draft campaigns back.
ProjectbetaPythonSite audit for SEO and AI-search visibility: a deterministic crawl, narrow Jev Choice, Score and Noul questions, and a live dashboard.
AppbetaTypeScriptChrome extension that grades readable page sections on ten rubric scores with Jev via Vercel AI Gateway and gives the page an A-E grade.
ProjectbetaTypeScriptWatches X for updates relevant to your YAML profile, has Jev classify and score each post, and alerts in Slack.
ProjectbetaTypeScript