Project overview
The problem we solved
Partnership and growth teams need creator intelligence at volume — not a handful of manually reviewed channels. The brief was clear: analyze up to 10,000 YouTube and TikTok channels per day for engagement quality and brand safety, without blowing the budget on cloud AI infrastructure.
The challenge
What teams were stuck with
Per-channel transcript and content analysis is expensive. Running premium models on every video at 10K channels/day was not viable, and managed cloud queues would erase the margin.
- Target throughput of 10,000 channels per day across YouTube and TikTok
- Engagement scoring and brand-safety had to come from real transcript analysis, not vanity metrics
- Strict cost ceiling ruled out naive full-depth LLM analysis on every channel
- Managed cloud workers and queues would dominate the infrastructure bill
- A single monolithic prompt could not reliably cover scoring, safety, and partnership signals
Our solution
What we built
Vyrex is a multi-agent creator analysis platform: specialized Claude agents handle engagement and brand-safety from transcripts, while tiered pipelines and self-hosted Redis + Celery on Hetzner keep cost per channel under control.
- Claude Agent SDK orchestrating 5+ specialized agents for scoring and brand-safety
- Transcript-based analysis for engagement quality and safety signals
- Tiered pipelines — cheap/fast passes first, deep analysis only when signals warrant it
- Self-hosted Redis + Celery on Hetzner for queue workers at low infra cost
- YouTube and TikTok coverage with partnership scores searchable in natural language
- Live product at vyrex.io for discovery, profiles, and export
Technology
Built with enterprise-grade technology
Goals and objectives
What the platform needed to deliver
01
Hit 10K channels/day
Sustain high daily throughput without queue collapse or runaway LLM spend.
02
Score engagement that matters
Move past views and subs — analyze transcripts for buying intent, trust, and real audience engagement.
03
Detect brand-safety risks
Flag unsafe or off-brand creators from content signals before budget is committed.
04
Keep infrastructure cheap
Self-host Redis and Celery on Hetzner instead of paying cloud premiums for background work.
05
Specialize with multi-agent design
Give each agent a clear job — scoring, safety, partnership signals — instead of one overloaded prompt.
06
Tier depth by value
Run light analysis at volume and reserve expensive deep passes for high-potential creators.
Solution in action
See the platform in action
From discovery to decision — how Vyrex analyzes creators at scale with multi-agent AI.
01
Landing — find creators before you waste budget
Vyrex positions influencer discovery around brand-fit, audience intent, and partnership signals — not vanity follower counts.
02
Partnership analysis & AI recommendation
Multi-agent output lands as actionable cards: audience intent, partnership signals, growth timing, and a clear SCALE / WAIT / SKIP recommendation.
03
Creator profile — brand fit & audience
Each profile shows best-fit categories, audience demographics, buying intent, sales likelihood, and whether brands actually return.
04
Semantic search results at scale
Natural-language queries return thousands of YouTube and TikTok creators with partnership scores, brand-safety tags, and CSV export.
05
AI semantic search
Teams describe ideal partners in plain English — VPN promoters, high buying intent, SaaS-safe niches — and the index returns ranked matches.
Platform architecture
How it all works together
1
Claude Agent SDK orchestration
5+ specialized agents collaborate on engagement scoring, brand-safety, and partnership signals instead of a single monolithic analysis.
2
Transcript analysis pipeline
Video transcripts feed agents that evaluate buying intent, trust, safety, and sponsor patterns from what creators actually say.
3
Tiered analysis depth
Lightweight passes run at volume; expensive deep analysis triggers only when early signals justify the cost.
4
Self-hosted Redis + Celery
Background jobs run on Hetzner-hosted Redis and Celery workers — high throughput without managed-queue premiums.
5
YouTube + TikTok ingestion
Channel metadata, content, and transcripts across both platforms feed a shared partnership scoring model.
6
Searchable creator index
Analyzed creators become discoverable via semantic and direct search with scores, safety tags, and export.
Results
Value and impact delivered
What changed once the platform was live.
Scalable creator analysis
Architecture supports the 10,000 channels/day target without collapsing into manual review.
Infrastructure costs kept low
Self-hosted Redis + Celery on Hetzner plus tiered pipelines keep cost per channel under the client's constraint.
Engagement beyond vanity metrics
Transcript-driven scoring surfaces buying intent and trust — not just views and subscriber counts.
Brand-safety before spend
Dedicated agents flag risk early so teams avoid creators who look big but are unsafe for the brand.
Decisions teams can act on
Profiles end in clear recommendations — SCALE, WAIT, or skip — with partnership and growth evidence attached.