How I Automated Social Media for My SaaS (and Got My First 10 Leads)
Three hours a day. That's how much time I was spending on social media before I snapped.
Scroll Twitter. Find a conversation. Think of something clever. Type it out. Delete it. Type something less clever but more honest. Post it. Move to LinkedIn. Repeat. Check if anyone responded to yesterday's stuff. They didn't. Start over.
I'm a senior engineer. Ex-Amazon, ex-CTO. I've been trying to build indie products for 5 years. Zero recurring revenue. And every time I'd dig into why, the answer was the same: nobody knows you exist. The product isn't the problem. Distribution is the problem.
So I kept doing the thing I hate most. Posting. Engaging. Pretending I enjoy the grind of "building in public" when what I actually wanted was to build the product.
The breaking point
It was a Tuesday. I'd spent 2.5 hours commenting on threads, got zero responses, and realized I had the same number of followers as the week before. I sat there staring at the screen and thought: I'm a systems engineer. Why am I doing this by hand?
That night I started building what became Galevox.
The first version was embarrassing. A cron job that ran claude -p with a prompt that said "find interesting tweets about developer tools and reply to them." It worked for about a day before it started posting generic garbage that sounded like every other AI reply bot.
But the bones were there. The idea that a system could find conversations, participate in them, and track what happens next.
What it does now
Eight months later, the system runs 20 autonomous agents across X and LinkedIn. It handles 90 sessions a day. 75 replies on X, comments on LinkedIn, original posts on both platforms.
The whole thing costs about $10/day all-in. Claude Max subscription, Apify for discovery, a small VPS for orchestration. That's the honest number.
Here's what surprised me: the first month was terrible. The replies were fine, technically. Grammatically correct, on topic, not obviously robotic. But they didn't convert. Nobody clicked through. Nobody DMed. 90,000 impressions in a month and maybe 2 warm leads.
The topic scoring revelation
The topics that get engagement are not the topics that get leads.
I built a scoring system that weights topics by DM conversion rate, not by impressions or likes. The formula looks at reply rate, recency, diversity, and most importantly, how often a topic actually leads to someone reaching out.
Turns out, broad tech conversations ("what's your favorite framework?") get tons of likes but zero leads. Specific pain point conversations ("I've been posting on LinkedIn for 6 months and nothing is working") get fewer likes but actual humans reaching out.
The system now runs a query intelligence pipeline every morning. It scores topics, retires what's stale, generates new search queries, and rotates its focus. All automatic.
The voice problem
Early on, people could tell my replies were AI-generated. Not because the grammar was wrong, but because the voice was wrong. Too polished. Too balanced. Real people have rough edges, tangents, opinions they state without hedging.
So I built a voice fingerprint system. It analyzed 50+ of my real posts, extracted the patterns (sentence rhythm, vocabulary, how I start and end thoughts), and injects that fingerprint into every session. There's even a drift detector that flags when the output starts sliding back toward generic AI tone.
It took about 3 months before the output stopped feeling like a template. That was the real breakthrough, not the automation itself, but making the automation invisible.
The 10 leads
They came from unexpected places. A founder building a meditation app who was frustrated with social media management. A marketing agency owner who couldn't scale content for 12 clients. A developer advocate who wanted to automate community engagement.
None of them came from my "best performing" posts. They came from replies in conversations where someone was venting about a real problem, and the system responded with something specific and useful.
The 3-layer memory system helped here. It tracks cycle debriefs (what happened each session), consolidates them into working memory (what's been working this week), and distills long-term learnings (what we know about our audience after months of data). Each session starts with that context, so the system doesn't just react to individual tweets. It remembers.
What I'd tell past me
Stop optimizing for impressions. Start optimizing for conversations with the right people.
Don't try to sound smart. Try to be useful.
And for the love of god, stop doing it manually. You're an engineer. Build the system.
If you're a solo founder or agency owner tired of the posting treadmill, book a 30-minute demo and see the system running live. Or get the playbook -- free PDF on how we run SMM for $10/day.