Case study · Applied AI
Governed RAG for SME market intelligence, in Spanish, with citations.
Continuous, evidenced insight into what SMEs and the self-employed discuss online: taxes, social security, e-invoicing, hiring, cash flow. Manual browsing did not scale and could not answer with sources. We built a pipeline where nothing reaches the knowledge base unreviewed and every answer is cited.
- Client
- Spanish market-intelligence venture(anon.)
- Industry
- Applied AI
- Duration
- Pipeline, app and handover
- Team
- Senior pod, CORTEX
The challenge
What stood in the way.
The venture needed continuous, evidenced insight into the problems SMEs and self-employed professionals discuss online. Manual browsing did not scale, left no record, and could not answer a question with sources. Any system had to be trusted by an operator before its output was trusted by a client.
Our approach
How we framed it.
Collect, score, de-duplicate, review, then answer. Seven adapters pull posts per topic every twelve hours. One router turns each post into a structured signal scored 0 to 15 across five components and meters token cost per call. Identity, text hash and vector similarity are checked before storage, with a dead-letter queue if a check cannot run. An operator approves or rejects every signal; nothing is stored unreviewed.
The orchestration was designed around n8n Cloud's 300-second execution ceiling with chunked, paced dispatch and self-continuation, Basic Auth on every webhook, HMAC sessions, encrypted settings, and one error-handler workflow every other workflow reports to.
The solution
What we built.
- 25 n8n Cloud workflows across 7 platforms via Apify scrapers and the YouTube Data API
- One LLM router (Gemini, Claude and OpenAI) scoring five components, token cost metered per call
- Three-layer de-duplication gate: identity, text hash, vector similarity
- Operator review queue before the knowledge base
- Spanish chat over Pinecone that answers only from approved, cited signals, or says the evidence is short
- Next.js and TypeScript app on MongoDB Atlas and Railway
- Handover: in-app help, a written manual and 27 narrated training videos
Results
What changed.
A twelve-hour intelligence cycle across seven platforms with a human approval gate, and a chat that answers in Spanish with citations or declines. The client's operators run it themselves.
Stack
Tools used in production.
- n8n Cloud
- Next.js
- TypeScript
- MongoDB Atlas
- Pinecone
- Railway
- Gemini
- Anthropic Claude
- OpenAI
- Apify
- YouTube Data API
Talk to us
Have a similar problem? Bring it to a senior engineer.
A senior engineer plus the relevant department lead joins the first call. No discovery gauntlet, no junior reps.