re-publisher

01 / The task

Telegram collection and content processing

Content goes through extraction, classification and draft preparation. The source and intermediate results are stored separately, so a failed step can be inspected and rerun.

Use case: Editing content for topic-based channels.

Work done

• Normalization, data extraction, classification and editing templates.
• Processing history and results for individual steps.
• Image build, migrations and deployment of the original interface. External integrations still need configuration.

02 / Technology

Stack

03 / Decisions

How the project works

• The original application's app.web.main:app is running, rather than app.portfolio.
• PostgreSQL with the original migrations; separate persistent sessions, media, models, cache, artifacts and reports directories.
• The collector and full-cycle pipeline are separate services in the connected profile. They are activated after configuring the Telegram session, model and selected integrations.
• The administrative interface is protected by HTTPS Basic Auth; secrets are kept outside the repository.

04 / Verification

Demo workflows

1. Sign in to the owner's interface and open Materials, Drafts, Labeling, Models and Pipeline.
2. Open Rewrite Progress, Prompts and Search: these are the original pages backed by PostgreSQL.
3. After configuring Telegram: app.main login/login-qr → inspect-folder → limited sync → inspect materials.
4. After registering a model and configuring the LLM: processing → classification → draft → review → approved publication to the selected channel.

Check results

• Five sequential HTTPS GET requests from the VPS to itself: 81.5–228.6 ms, all HTTP 200. This is not a load test or an assessment of the external network.
• No Telegram collection, LLM processing or Codex task ran during the measurements.

Limitations

• Telegram API ID/hash and an authorized MTProto session have not yet been provided. Collection of real messages has not started.
• The active tfidf_logreg.joblib model and selected LLM settings are missing. There is no basis for claiming verified classification or rewriting quality.
• The interface currently shows an empty production-schema database. The previous demo's synthetic corpus has not been mixed into the original tables.
• External publishing has not been tested: it requires a destination channel and BOT_TOKEN or MAX settings. AUTO_PUBLISH remains false until the destination is configured.