Cortex: An AI Enrichment Engine for Creative Assets
Cortex enriches every creative asset the moment it's produced transcribing, analyzing, embedding and tagging automatically, so cost and search are observable in real time, not estimated after the fact.
$0.03
Per-Creative AI Analysis Cost
3,869
Creatives Enriched to Date
80:1 – 240:1
Labor-to-AI-Spend Ratio
0.46 – 1.47
Full-Time Equivalents Saved / Year
The Challenge
High-volume creative production has an output problem that turns into a data problem assets land as opaque files with no summary, tags, hooks or audience signal.
No Search
You cannot find "the video where the spokesperson makes a bold claim" unless someone remembered to label it.
No Taxonomy
Hooks, angles and ideal-customer profiles live in people's heads, not on the assets.
No Analytics
Without structured fields there is nothing to chart, compare or optimize against.
No Strategy Reuse
Personas and briefs are rebuilt by hand every campaign — slow, inconsistent, expensive.
Our Solution
Cortex is an asynchronous, webhook-driven enrichment service, the heavy work runs in background tasks so the caller's request returns in milliseconds.
Multimodal Creative Analysis
Gemini reads the image or video and returns a summary, hook, angles, personas, tone and keywords.
In-House Transcription
Deepgram Nova-3 builds a timestamped transcript that feeds the analysis as ASR ground truth.
Semantic & Hybrid Search
Every creative gets a vector embedding, making the library searchable by meaning, not filename.
Taxonomy Embedding & Matching
A cosine prefilter plus LLM re-rank matches each creative to the account's tags and hooks.
Business Impact
Every AI call is priced the instant it happens, so cost is observable per creative and thousands of hours of manual tagging become an automatic property of ingestion.
$9,500 – $30,400
Net financial benefit modeled across all 7 capabilities
963 – 3,052 hrs
Human hours avoided per year
5–15 sec
Persona, brief and portrait turnaround, vs. 45 min–3 hrs by hand
Searchable Library
Assets found by meaning; personas and briefs generated on demand, not rebuilt
Module Overview
A look inside module interface.
Dashboard
The Analysis Pipeline
The lightweight, secure data pipeline.
Ingest
The host app calls Cortex; the request returns in milliseconds via 202 Accepted.
Transcribe
Deepgram Nova-3 transcribes video audio into timestamped segments.
Analyze
Gemini's multimodal model generates the summary, hook, angles and personas.
Embed & Match
A vector embedding is generated and matched against the account's taxonomy.
Notify
The structured record saves to MongoDB and a signed webhook notifies the caller.
Evolvan's Role
Evolvan designed and built Cortex end to end the FastAPI microservice, the analysis pipeline, the cost-tracking ledger, and the production safeguards that keep it auditable and unattended.
Analysis Pipeline Engineering
Evolvan built the fixed-sequence pipeline — download, transcribe, analyze, embed, match, notify — with a cosine prefilter and a deterministic LLM re-rank fallback.
Cost-Telemetry System
Evolvan built the per-call pricing ledger — a versioned rate sheet and a fire-and-forget recorder that prices every AI call the instant it happens.
Production Hardening
Evolvan added SSRF defense on outbound fetches, prompt-injection boundaries, scrubbed error surfaces, and end-to-end OpenTelemetry tracing.
Bulk & Async Infrastructure
Evolvan engineered the idempotent, resumable, cancellable bulk-analysis run and the signed-webhook delivery model.