JevPulse

High-throughput qualitative consensus engine for YouTube comments. Evaluates every comment individually with calibrated decision models.

JevPulse replaces conversational summarization with a Dual-Process AI Architecture: use a reasoning model once to design the schema, evaluate thousands of comments in parallel using TypeSafe Jev, and compute consensus deterministically in code.

Live Demo: jevpulse.vercel.app

System Architecture

The pipeline decouples high-level reasoning from high-throughput execution across four stages:

---

Architectural Stages

Stage 1: Dynamic Rubric Synthesis (System Two)

Static categories (positive, negative, neutral) fail across different domains. A coding tutorial needs evaluation on code clarity and pacing, while a hardware review needs evaluation on thermal throttling and pricing value.

Gemini Flash is invoked once per video using the title, description, transcript excerpt, and 20 sample comments. It outputs a structured rubric where each criterion defines:

• Evaluative Question: Closed question directed at an individual comment.
• Affirmative Boundary: Explicit criteria for what fulfills the dimension.
• Negative Distractor Boundary: Similar-sounding remarks that must be rejected to prevent semantic drift.

Stage 2: Dual-Layer Question Topology (System One)

For every comment in a batch, the engine builds a two-layer evaluation graph:

• Layer 1: Universal Semantic Invariants (Runs on every video)
  • is_opinion (noul): Binary probability separating substantive discourse from greetings, spam, and timestamps.
  • comment_type (choice): Categorical intent (praise, criticism, question, suggestion, agreement, disagreement, correction, experience, other).
  • specificity (score): 3-point ordinal score measuring argument depth and empirical detail.
  • has_claim (noul): Binary probability measuring whether the utterance asserts a verifiable factual claim.
• Layer 2: Dynamic Rubric Stance (Video-Specific)
  • One noul question per criterion generated in Stage 1, measuring the comment's affirmative probability on that specific dimension.

Stage 3: High-Throughput Speculative Fan-Out

Comments in each batch are bundled into a single shared state. Jev ingests the state once and evaluates all Layer 1 and Layer 2 questions across all batch comments concurrently in memory. Adding questions barely impacts latency because decisions evaluate in parallel over shared memory rather than generating tokens sequentially.

Stage 4: Deterministic Synthesis & Consensus Math

Jev outputs typed numbers and distributions. Application code retains complete ownership of aggregation, thresholding, and business logic:

• Signal Gating: Filters out low-entropy comments using calibrated isOpinion and specificity thresholds before computing consensus.
• Consensus Ratios: Computes true percentage agreement across the filtered substantive manifold.
• Principled Dissent Isolation: Silence is not dissent. A comment is classified as opposing only if it has a low affirmative score and actively exhibits critical communicative intent (criticism, disagreement, correction).
• Confidence-Calibrated Provenance: Ranks supporting and dissenting evidence by absolute semantic confidence ($|P - 0.5|$) rather than social likes, preventing viral jokes or spam from distorting the report. Every metric directly links back to source comment IDs.

---

Performance & Cost

Metric	Traditional Chat LLM (GPT-4o / Claude 3.5)	JevPulse (Gemini + TypeSafe Jev)
Mechanism	Autoregressive text generation	Calibrated discriminative decisions
Input Pricing	~$2.50 – $3.00 / Mtok	$0.042 / Mtok (outputs free)
Cost per 3,000 comments	~$2.40 – $4.50	~$0.025 – $0.04 (~100x cheaper)
Total Evaluation Latency	45 – 120 seconds	~1.6 – 2.0 seconds
Schema Guarantees	Probabilistic JSON parsing errors	Strictly typed values and floats
Data Integrity	Compression loss / Hallucinated counts	100% individual comment provenance

---

API Endpoints

• POST /analyze: Analyzes a YouTube video URL and returns the complete consensus report as JSON.
• GET /analyze/stream: Server-Sent Events (SSE) endpoint providing real-time batch progress, processed comment counts, and live decision streaming.

---

Getting Started

Prerequisites

• Node.js 18+
• API keys for TypeSafe AI, Google Gemini, and YouTube Data API v3

Installation

# Clone the repository
git clone https://github.com/jaygajera17/JevTube.git
cd JevTube

# Configure environment variables
cp .env.example .env

# Install dependencies
npm install

# Start local server
npm run dev