AI-Powered LinkedIn Engagement Assistant

Finding the right conversations to join — without manually searching through LinkedIn every day

Tools: n8n • Tavily API • Google Gemini • JavaScript • LinkedIn Public Posts
Focus: AI Automation • Business Process Automation • Professional Networking • AI-Assisted Decision Making

The Problem

Building a professional network on LinkedIn isn't only about publishing content. Meaningful participation in relevant conversations can also help build visibility, relationships, and domain credibility.

But there is a practical problem:

Finding the right posts to engage with takes time.

A manual process typically involves searching LinkedIn, scrolling through posts, checking how recent they are, reading multiple posts, deciding whether they are relevant, thinking about whether I can genuinely contribute something useful, and finally deciding what to comment.

Most posts discovered during this process may not actually be worth engaging with.

So I asked:

Can AI automate the discovery and evaluation process while keeping the actual engagement human?

That became the objective of this project.

The Solution

I built an AI-powered LinkedIn Engagement Assistant using n8n.

Instead of automatically liking or commenting on posts, the workflow focuses on the more time-consuming research process:

discover → filter → understand → evaluate → rank → recommend

The workflow searches for recent public LinkedIn posts related to areas relevant to my professional interests, including:

• Business Analytics

• Revenue Operations

• Sales Operations

• GTM

• Sales & Marketing Analytics

• Revenue Forecasting

• Power BI

• AI Automation

• Business Process Automation

It then evaluates the posts and returns a shortlist of the Top 5 engagement opportunities, together with an explanation and a suggested comment.

The final decision — including whether to comment and what to actually write — remains with me.

How the Workflow Works

1. Generate Targeted Search Queries

The workflow starts by generating predefined searches covering the professional areas I want to follow.

Instead of relying entirely on the LinkedIn feed, this creates a focused discovery process around topics that are relevant to me.

2. Discover Public LinkedIn Posts

The workflow uses the Tavily Search API to discover publicly accessible LinkedIn post URLs matching those searches.

Search results from multiple topics are combined into a single candidate pool.

3. Clean and Deduplicate Results

The workflow removes irrelevant URLs and duplicate posts.

This is particularly important because the same LinkedIn post may appear in results for multiple search queries.

4. Verify the Actual Age of the Post

Search-engine freshness alone isn't enough.

For compatible LinkedIn post URLs, the workflow extracts the LinkedIn activity ID and derives the post timestamp.

Posts outside the required freshness window are removed.

For the current version, I focus on posts published within approximately 24 hours.

This matters because the value of joining a professional discussion often decreases once the conversation is already old.

5. AI Preliminary Screening

The remaining candidates go through an initial Google Gemini screening.

At this stage, AI classifies each post as:

KEEP — MAYBE — DROP

Obvious low-value engagement opportunities such as job advertisements, appointment announcements, promotional content, congratulations posts, unrelated material, and search noise can be removed.

Importantly, the model isn't instructed to select a post simply because it contains keywords such as “RevOps” or “AI.”

It evaluates whether there appears to be an actual business discussion where I could contribute something useful.

6. Extract More Complete Post Content

Search snippets don't always contain enough information to properly evaluate a LinkedIn post.

The shortlisted URLs therefore go through Tavily Extract to retrieve more of the publicly available post content.

The workflow then cleans unnecessary page content and LinkedIn boilerplate before deeper analysis.

7. Deep AI Scoring

Gemini performs a second, more detailed evaluation.

Each candidate is scored across four dimensions:

Dimension

Weight

Professional relevance

30

Opportunity to add value

30

Networking value

20

Discussion quality

20

Total

100

This is an important distinction.

The system isn't asking:

“Does this post contain one of my keywords?”

It is asking:

“Is this a conversation where my professional background could allow me to contribute something meaningful?”

8. Select the Top 5

The posts are ranked according to their overall opportunity score.

Only the five strongest opportunities are presented in the final output.

For each recommendation, I receive information such as:

Rank • Opportunity Score • Post Age • Post Title • LinkedIn URL • Why It Was Selected • Suggested Contribution Direction

9. Generate a Suggested Comment

For each selected post, AI can also propose one thoughtful comment based on the post and the identified contribution opportunity.

The prompt is specifically designed to avoid fabricating personal experience.

For example, rather than falsely saying:

“I've seen sales teams struggle with this…”

when that specific experience hasn't been established, the system can frame the contribution analytically:

“One interesting extension would be to build this check into the CRM itself…”

This allows AI to assist with thinking without pretending to have experiences on my behalf.

10. Human Review and Manual Engagement

The workflow intentionally stops before publishing anything.

I open the original LinkedIn post, read it myself, review the AI recommendation, modify or reject the suggested comment if necessary, and manually decide whether to participate.

So the final architecture is:

AI for discovery and decision support. Human judgment for actual engagement.

Example

One of the posts discovered by the workflow discussed improving revenue forecasts by requiring every committed opportunity to have a confirmed next step and confirmed date.

The system ranked the post highly because it connected strongly with CRM, Sales Operations, pipeline management, forecasting and data quality.

It then suggested an additional angle: instead of relying only on manual pipeline audits, this rule could potentially be incorporated into CRM validation.

A generated comment was:

The "confirmed next step" framework is excellent for grounding forecasts in reality. To scale this, teams could automate a CRM validation check that flags opportunities with missing calendar events or expired "next step" dates. Moving these data quality checks from manual audits to automated alerts can help maintain forecast hygiene consistently across the entire pipeline.

The value here isn't merely that AI wrote a comment.

The more interesting part is that the system found the conversation, determined that it was relevant, identified why I could contribute, ranked it against other opportunities, and proposed a potential contribution.

Why Not Just Search LinkedIn Manually?

Manual searching works.

The problem is the amount of low-value work surrounding the few useful conversations.

Suppose dozens of candidate posts are discovered across several topics.

Manually, I would need to open or inspect many of them, determine their freshness, remove duplicates, identify promotional or irrelevant content, read the useful ones, compare their relevance, and decide where my time is best spent.

The automated workflow performs much of that filtering first.

Instead of asking:

“What should I search for today?”

I can start with:

“Here are today's five strongest opportunities. Which ones are actually worth my attention?”

The objective isn't to eliminate human judgment.

It's to spend human attention where it has the highest value.

Why This Approach Is Useful

There are three aspects of the project that I find particularly important.

First, AI is being used for prioritization rather than indiscriminate content generation. The system evaluates many possible opportunities and helps narrow them down.

Second, the workflow uses multiple levels of evaluation. Cheap initial discovery and filtering happen before more detailed extraction and AI analysis. More expensive reasoning is therefore concentrated on a smaller candidate set.

Third, it remains human-in-the-loop. The system can recommend a post and suggest a response, but it doesn't decide what gets published under my identity.

What Can Be Extended Further?

The current version deliberately focuses on creating a useful working system rather than automating every possible step.

The workflow can later be extended to:

• Export recommendations to Google Sheets

• Maintain a history of previously discovered posts

• Prevent repeated recommendations

• Track posts I actually engaged with

• Analyze which topics consistently produce useful conversations

• Track recurring authors

• Deliver recommendations through email or messaging

• Add additional ranking signals

• Improve extraction of different LinkedIn content formats

• Create engagement analytics over time

For example, a Google Sheets integration could store:

Date | Post | Author | URL | Topic | Score | Recommendation | Suggested Comment | Engagement Status

This would turn the workflow from a daily discovery tool into a longer-term engagement intelligence system.

Current Limitation

The workflow primarily works with publicly accessible text-based LinkedIn posts that can be discovered and extracted through search.

It may not reliably understand image-only posts, carousels, videos, private/restricted content, or posts that aren't adequately indexed.

I consider that an acceptable limitation for the first version because the objective was not complete LinkedIn coverage.

The objective was to prove that a significant portion of the discovery → filtering → prioritization → recommendation process could be automated effectively.

What I Learned

This project reinforced an important lesson about practical AI automation:

The best automation isn't necessarily the one that removes the human. Sometimes it's the one that removes low-value work before the human needs to make a decision.

The most valuable part of this workflow isn't automatic commenting.

It's reducing the effort required to discover the few conversations where my time and knowledge may actually be useful.

Project Stack

Workflow Orchestration: n8n
Web Search & Extraction: Tavily API
AI Evaluation: Google Gemini
Custom Processing: JavaScript
Source: Publicly discoverable LinkedIn posts
Optional Storage/Reporting: Google Sheets
Final Decision: Human review