Technical Reference
How Rally works under the hood — the data pipeline, AI models, and automation that power your engagement workflow.
Pipeline Architecture
Rally runs a four-stage AI pipeline to find, analyze, and prepare engagement opportunities.
Ingest
Apify
Analyze
Gemini
Score
Claude
Surface
Claude
Ingest
Apify
Analyze
Gemini
Score
Claude
Surface
Claude
AI Prompt Chain
A chain of specialized prompts, each building on the previous output.
Gemini 2.0 Flash analyzes the actual video — visual content, audio, captions, mood, brand safety, and suggested comment angles.
Claude scores relevance (0-100), timing (0-100), and reach (0-100), producing a composite confidence score with reasoning.
Claude generates 3 comment variants using your brand voice, video context, and scoring insights. Each has a distinct style.
Self-critique loop — checks comments against voice guidelines, banned words, and quality standards. Regenerates up to 2x.
Scoring Model
The composite score (0-100) is the primary ranking metric.
Topic match, audience overlap, hashtag alignment
Hours since posted, comment window opportunity
Views, likes, followers, growth velocity
Positive content preferred
High controversy = lower score
Unsafe = excluded entirely
Posts scoring below 65 are filtered out by default. Threshold is configurable per brand.
Automation Schedule
Background jobs run on scheduled intervals to keep your queue fresh.
Content fetch
Pull fresh TikTok content via Apify
Video analysis
AI analysis + scoring of fetched videos
Agent cycle
Observe/decide/act loop for surfacing
Learning pass
Process user feedback to improve scoring
Workmate report
Generate daily summary and approval queue
Feedback Loop
Every action you take teaches Rally what matters to your brand.
Reinforce this post type + voice style
Deprioritize this topic or creator
Adjust voice and tone calibration
Quality bar not met — improve generation
Rally uses the last 50 actions to build learned preferences via the curation-analysis-with-feedback prompt variant.