← All projects

2026 · Agentic content pipeline

LinkedIn Post Swarm

A multi-agent AI pipeline that researches, drafts, critiques, and publishes LinkedIn posts — with a human approval gate before every post.

Live 2026

Why I built this

Maintaining a consistent LinkedIn presence was taking hours I did not have, and I kept skipping it. The system had to match my voice exactly — no hustle content, no vague motivation, only genuine technical signal. The harder engineering problem was keeping the human in the loop without making the approval flow annoying enough to skip.

~1/day Follower requests (was ~1/month)
Engagement growth in 1 month
6 Posts published

Architecture

Research Agent

Qwen2.5 14B Q4 via Ollama. Fetches and relevance-scores stories from 7 RSS and web sources with CoT reasoning. Prefers today's news with a 3-day fallback. Returns top-3 story briefs to the orchestrator.

Claude Orchestrator

claude-sonnet-4 via API. Picks story, writes post in persona, manages the revision loop, and escalates to Telegram if the critic loop exceeds 10 iterations. Prompt caching on all static context blocks to reduce API cost.

Critic Agent

Qwen2.5 14B Q4. Scores drafts on 4 weighted axes: authenticity (30%), depth (25%), recruiter signal (20%), hook quality (15%). Threshold 6.5. Returns structured JSON feedback with improvement directions for each revision.

Telegram Approval Bot

9 action buttons per draft: Approve, Regenerate (with optional direction text), Reject, Edit inline, Undo (3-step history), As Carousel (PDF), As Article, With Image, and cross-post toggles for Twitter/X and Bluesky.

Playwright Publisher

Persistent browser session posts at scheduled time (Mon/Thu 8:00 AM IST). Screenshots confirmation. Supports Text, Carousel PDF (fpdf2), LinkedIn Article, and image-attached formats.

Journal System — Monday posts

LLMLingua-2 compresses journal.md entries (≥60% token reduction). Mistral 7B extracts a structured brief. Claude writes personal-update posts from the compressed work log, giving Monday posts a different voice from Thursday external-discovery posts.

Tech stack

Technologies used

core

Claude claude-sonnet-4 (orchestration)Qwen2.5 14B Q4 (research + critic)Mistral 7B (scheduler)Ollama (local SLM hosting)

infra

Playwright (browser automation)Telegram Bot API (approval)FastAPI + HTMX (dashboard)RTX 3080 Ti (12 GB VRAM)

tools

fpdf2 (carousel PDFs)tweepy (Twitter/X)atproto (Bluesky)LLMLingua-2 (compression)RSS feed parser

Key highlights

Proof points

  1. 01

    LinkedIn follower requests grew from ~1/month to ~1/day within one month of deployment.

  2. 02

    Post engagement grew from ~400 to ~2,000 impressions per month in four weeks.

  3. 03

    Zero autonomous posts — every draft routes through Telegram approval with full edit, undo, and regenerate controls.

  4. 04

    Critic agent enforces a 6.5/10 quality threshold across 4 axes before any draft reaches the approval queue.

  5. 05

    6 posts published; cross-posting support live for Twitter/X and Bluesky.

Focus areas

Agent orchestrationPrompt engineeringPlaywrightTelegram Bot APIWorkflow reliability

Explore the work