Built at work · description only

SemanticAI Briefing Suite

Keyword in, semantic content brief out. A multi-step pipeline, not a single prompt, with a human approval step before anything is used.

Problem

SEO content briefs were written by hand. Research ran across several tools, one brief took up to an hour, and there was no shared format. Search had become semantic. The briefs had not.

Solution

A web app that runs each brief through a fixed chain of steps:

  1. Keyword and campaign details as input
  2. SERP and competitor analysis
  3. Embeddings as the semantic core of the analysis
  4. Gap detection and open questions
  5. Structure and meta proposal
  6. Review loop: the generated brief is checked against the keyword and the SERP data
  7. Review and approval by a person
  8. Archive

An LLM handles steps 4 to 6. Progress is streamed to the browser while the steps run. Access is protected by a login.

Architecture of the SemanticAI Briefing Suite A web app with login streams progress via server-sent events. Eight steps run in order: keyword and campaign input; SERP and competitor analysis using a SERP API; embeddings using an embedding model; gaps and open questions, structure and meta proposal, and a review loop that checks the brief against keyword and SERP data, all three handled by an LLM; human review and approval; and an archive in a database. Web app Login · live progress in the browser via SSE 1 Keyword + campaign info 2 SERP + competitor analysis 3 Embeddings: semantic core 4 Gaps + open questions 5 Structure + meta proposal 6 Review loop brief vs. keyword + SERP data 7 Human review + approval 8 Archive SERP API Embedding model LLM Database
Schematic overview. No internal interfaces or data shown.

Stack

  • Python
  • FastAPI
  • Pydantic
  • PostgreSQL
  • Server-Sent Events
  • OpenAI text-embedding-3-large
  • Claude with GPT as fallback
  • serper.dev

Result

6 to 10 minutes instead of up to an hour per brief.

One consistent format for every brief. In daily use by the entire SEO team since July 2026.

My role

I designed, built and rolled out the tool on my own.

What I learned

One prompt is not enough. The human approval step was the most important design decision.

Back to all projects