Skip to main content
    The intelligence stack

    The pipeline behind every brief.

    No black boxes. Every source, every model, every weight — documented.

    2.1M
    Videos in corpus
    70+
    Countries indexed
    <60s
    Brief latency p95
    94%
    Breakout precision
    Pipeline

    Six stages, end to end.

    From raw API call to streamed brief, in the order it actually runs.

    STAGE 01

    Ingest

    Pull deltas from YouTube + Trends across 50 regions and 70 countries. Cached behind a Redis-backed quota router.

    STAGE 02

    Normalize

    Parse durations, dedupe by canonical video id, recompute velocity per channel baseline, attach geo + category metadata.

    STAGE 03

    Score

    Weighted blend (velocity 35 / lift 25 / novelty 20 / fit 20) into a 0–100 Virality Score with a confidence band.

    STAGE 04

    Cluster

    Embed titles + comments, run HDBSCAN to collapse near-duplicate angles, surface the dominant frame per cluster.

    STAGE 05

    Reason

    Route the cluster to the right LLM (see model matrix) for hooks, titles, thumbnail brief and script outline.

    STAGE 06

    Serve

    Stream the brief to the client with SSE; cache for 60 minutes keyed by region + niche + channel id.

    Data sources

    Every input, every cadence.

    What we pull, how often, from where, and what it carries.

    YouTube Data API v3
    Cadence · Every 60 min
    Coverage · 50 regions · all categories
    Payload · Search, trending, channels, video stats, comments
    Google Trends
    Cadence · Every 15 min
    Coverage · 70+ countries · regional + DMA
    Payload · Real-time interest, related queries, breakouts
    Comment corpus
    Cadence · Per deep-dive
    Coverage · Top 500 comments / video
    Payload · Embedding-clustered into intent buckets
    Channel telemetry
    Cadence · On connect + daily
    Coverage · Your last 90 days
    Payload · Watch-through, CTR baselines, tone snippet
    The score

    How the Virality Score is calculated.

    One number, four weighted inputs, transparent math. No magic.

    score = 0.35·velocity + 0.25·lift + 0.20·novelty + 0.20·fit
    VelocityViews/hour vs channel + niche baseline
    35%
    Search-intent liftGoogle Trends acceleration over 14d
    25%
    NoveltyCosine distance from existing corpus
    20%
    Channel fitMatch to your niches and tone vector
    20%
    Urgency tiers:
    75+ Watch
    85+ Act
    90+ Now
    Model routing

    The right model for the job.

    We don't pick one model and live with its tradeoffs. We route per task.

    Topic scoring
    Gemini 2.5 Flash
    Cheap, fast, JSON-strict — runs on every refresh.
    Deep-dive reasoning
    GPT-5
    Long-context synthesis across 30+ corpus videos.
    Hooks & titles
    Claude Sonnet
    Best-in-class for short-form copy with constraints.
    Thumbnail brief
    Gemini 2.5 Pro
    Multimodal grounding on the corpus thumbnails.
    Routing layer is abstracted — we swap models as better ones ship.
    Guarantees

    What we won't do.

    No scraping

    Every signal comes through an official API. Rate-limit aware, quota-budgeted, with a fallback key per source.

    Your data is yours

    Channel telemetry stays scoped to your workspace. We never train shared models on your transcripts.

    Multi-region edge

    Edge functions on Cloudflare, primary store in EU. p95 brief latency under 60s end-to-end.

    Methodology FAQ

    Questions about the data.

    Ready to ship trend-driven videos?

    Try TrendsIQ free — no credit card, no API keys, ready in 60 seconds.

    Start Free Talk to us