Dixta — AI Content Automation

Tracking environmental and regulatory sources by hand scatters editorial work across searching, selecting, writing and distributing.

A professional platform built at GreenSoft together with another developer, independent from the WhatsApp assistant. A Laravel core orchestrates the pipeline and a Python microservice handles extraction with Scrapling.

Turn configured sources into reviewable drafts while keeping human editorial control and pipeline traceability.

• Built the first version of the platform and a large part of its evolution: scraping, candidate selection, draft generation and publishing.
• Hardened the Python extraction microservice: article validation, URL-based deduplication and browser rendering when a site requires it.
• Implemented draft validations and the review flow: nothing is published without a person approving it.
• Migrated the draft generator model to DeepSeek.

• Laravel owns the database, queues, candidate selection, drafts and publishing; the FastAPI + Scrapling scraper is stateless.
• The normalized URL and its hash are the deduplication contract between both services.
• The generator model went through several stages: it started with local models and today it is DeepSeek. The Bitrix24 integration and the llama.cpp server were removed.
• The model must return a JSON object with fixed fields; if the result fails validation, the draft falls back to a deterministic template.
• Publishing is a separate action from approving: a draft goes through human review before it goes out.

• Daily extraction from configured sources, with relevance and age filters.
• Article persistence and deduplication before spending inference.
• Draft generation with DeepSeek, validated and normalized before being shown.
• A panel to edit, regenerate, approve or reject each draft, with LinkedIn publishing only after approval.

• Keeping extraction reliable across heterogeneous sources without coupling it to the database.
• Avoiding duplicates and irrelevant content before spending inference or editorial time.
• Separating automation from publishing: the model proposes and a person keeps the final decision.

• Clear boundaries between orchestrator and scraper make each part easier to test, replace and deploy.
• Draft quality depends as much on selection and context as on the model used.