Industrial AI Content Factory

Autonomous daily LinkedIn content pipeline for Arjun Acharya — Port Automation & AI Engineer at Prosertek (Bilbao, Spain). The system researches maritime and industrial automation news, avoids near-duplicate posts via vector memory, generates PhD-level long-form posts with Gemini 1.5 Pro, and publishes via the official LinkedIn REST API v2 (/v2/ugcPosts) using an OAuth 2.0 access token (w_member_social), unless dry-run mode is enabled.

Architecture

┌─────────────────────┐     ┌──────────────────┐     ┌─────────────────┐
│ post_calendar.csv   │────▶│ Orchestrator     │────▶│ PostGenerator   │
│ (editorial calendar)│     │ (daily pipeline) │     │ (Gemini)        │
└─────────────────────┘     └────────┬─────────┘     └────────▲────────┘
                                     │                        │
┌─────────────────────┐              │              ┌──────────┴────────┐
│ MarketResearcher    │──────────────┼──────────────▶│ Prompt w/ news   │
│ (RSS + site scrape) │              │              └──────────────────┘
└─────────────────────┘              │
                                     ▼
                            ┌─────────────────┐
                            │ RAGEngine       │
                            │ (ChromaDB)      │
                            │ dedupe + recall │
                            └────────┬────────┘
                                     │
                                     ▼
                            ┌─────────────────┐
                            │ LinkedInPublisher│
                            │ (OAuth /v2)     │
                            └─────────────────┘
                                     │
                                     ▼
                            ┌─────────────────┐
                            │ PostScheduler    │
                            │ (APScheduler)    │
                            └─────────────────┘

Installation

python -m venv .venv
.venv\Scripts\activate          # Windows
# source .venv/bin/activate     # Linux / macOS
pip install -r requirements.txt

Copy .env.example to .env and fill in API keys and credentials.

The .env file is gitignored and will not be committed or pushed; only .env.example (placeholders) belongs in the repository. If you ever ran git add .env by mistake, run git rm --cached .env and keep secrets only locally.

Configuration

Variable	Purpose
GEMINI_API_KEY	Google AI Studio key for Gemini
LINKEDIN_ACCESS_TOKEN	OAuth 2.0 access token with w_member_social (UGC Posts)
LINKEDIN_CLIENT_ID / LINKEDIN_CLIENT_SECRET	Only needed to refresh the token locally (see script below)
LINKEDIN_EMAIL / LINKEDIN_PASSWORD	Optional legacy placeholders (not used for publishing)
CHROMA_PERSIST_DIR	Persistent ChromaDB directory
POST_CALENDAR_PATH	CSV calendar (data/post_calendar.csv)
SCHEDULE_HOUR / SCHEDULE_MINUTE	Local cron time
TIMEZONE	IANA zone (default Europe/Madrid)
CALENDAR_SEQUENCE_START	Optional YYYY-MM-DD: day 1 of the CSV is this date; each following day uses the next row (wraps at N = number of CSV rows). If unset, the row follows day-of-year mod N (not run count). In GitHub Actions, set repo variable CALENDAR_SEQUENCE_START.
DRY_RUN	true to skip publishing

Test LinkedIn (no Gemini key)

With LINKEDIN_ACCESS_TOKEN set in .env, verify the token against LinkedIn (GET /v2/userinfo, then GET /v2/me if needed):

python main.py --test-linkedin

This does not publish anything. The access token must include OpenID scopes for userinfo (openid, profile, …) and w_member_social for posting. If you only authorized w_member_social, userinfo returns 401 — re-run the local OAuth helper below.

Refresh LINKEDIN_ACCESS_TOKEN locally (browser + terminal)

1. In the Developer Portal, open your app → Products and ensure Sign In with LinkedIn using OpenID Connect and Share on LinkedIn are added.
2. Under Auth, add this Authorized redirect URL (exactly):
   http://127.0.0.1:8765/oauth/callback
3. Put LINKEDIN_CLIENT_ID and LINKEDIN_CLIENT_SECRET in .env.
4. Run:

python scripts/linkedin_oauth_local.py

Sign in when the browser opens, click Allow, and the script will exchange the code and update LINKEDIN_ACCESS_TOKEN in .env.

Usage

# One-off run for “today’s” slot (1–100 cycle from day-of-year)
python main.py --run-now

# Specific calendar day (1–N)
python main.py --day 42

# Daemon scheduler (cron at configured local time)
python main.py --schedule

# No LinkedIn post — log only
python main.py --run-now --dry-run

Repository layout

professional_linkedin/
├── main.py
├── requirements.txt
├── .env.example
├── data/
│   └── post_calendar.csv
├── src/
│   ├── config.py
│   ├── post_generator.py
│   ├── market_researcher.py
│   ├── rag_engine.py
│   ├── linkedin_publisher.py
│   ├── scheduler.py
│   └── orchestrator.py
├── scripts/
│   └── linkedin_oauth_local.py
└── tests/
    ├── test_post_generator.py
    └── test_full_pipeline.py

Tech stack

• Python 3.10+
• Gemini AI (google-generativeai) for long-form posts
• ChromaDB for embeddings, post history, and market insight recall
• APScheduler with cron triggers and timezone support
• LinkedIn REST API v2 (ugcPosts, OAuth2 bearer token) via requests
• BeautifulSoup + requests for lightweight research scraping and Google News RSS
• Azure themes appear in prompts and calendar topics (cloud IoT patterns)

Tests

Unit + generator tests:

python -m pytest tests/ -v

Full pipeline (orchestrator) without real Gemini or LinkedIn:

python -m pytest tests/test_full_pipeline.py -v

These tests mock PostGenerator.generate_post and stub market research so Chroma + dry-run publish run end-to-end.

Live smoke test (real Gemini + HTTP research + Chroma; LinkedIn skipped if --dry-run):

# Requires a valid GEMINI_API_KEY and LinkedIn credentials in .env
python main.py --run-now --dry-run --day 1

If you see API_KEY_INVALID, create a new key in Google AI Studio and set GEMINI_API_KEY in .env.

License

MIT