🚀 Built an Open-Source AI Outreach Engine to Find Customers, Jobs & Investors ⭐
I got tired of spending more time finding the right people than actually talking to them.
While building Observer, my workflow looked something like this:
LinkedIn → Apollo → AI writing → spreadsheets → repeat.
All that just to answer one question:
Who should I talk to next?
Customers.
Investors.
Hiring.
Job opportunities.
Partnerships.
The workflow never really changes.
So I built an open-source AI Outreach Engine to automate it.
It can:
- Discover companies that match your criteria.
- Find the right people inside those companies.
- Verify and enrich contact information.
- Research prospects automatically.
- Generate personalized outreach.
The goal wasn't to build "another AI SDR."
It was to create a transparent, hackable workflow that founders and engineers can self-host, customize, and extend instead of paying for an expensive stack of tools.
If you're building something and founder-led outreach is becoming a bottleneck, I'd love your feedback.
The repo is open source, so feel free to roast it, contribute to it, or just steal ideas from it.
GitHub: https://github.com/rohitmalhotra1420/ai-outreach-engine
Bounce rate, for sure — and from an angle I didn't expect. We don't generate full AI emails (kept it templated + one manually-personalized line, mostly for the same "still sounds like a person" reason you're solving for), but we scrape contact info from real business websites at scale. Ran an MX check before sending and caught a scraped address that was syntactically perfect but had literally no mail server behind it — an artifact from a broken contact page, not a typo. Made me realize "looks like a valid email" and "is actually deliverable" are two completely different checks — same lesson as your verify-before-send point, just from the address-sourcing side instead of the content side.
The “who should I talk to next?” framing is useful. Most outreach tools focus on sending volume, but the real bottleneck is usually picking the right people in the first place.
How are you handling the quality of the generated messages once volume goes up? That’s usually where these systems start feeling generic.
You could check out here:
https://github.com/rohitmalhotra1420/ai-outreach-engine
Most system generate geenirc mail during icps are generted. I do a dedicated LLM call with all ICP info and then get a email template.
Its editable for all/ every single contact as well.
Got it ,doing a dedicated LLM call with the full ICP info instead of a generic template is a smarter approach. Being able to edit per contact also helps keep it from feeling robotic.
Thanks for the GitHub link, I’ll take a look. Interesting project.
This hits close to home — I've been building something in the same space (automated prospecting/outreach). The scraping/discovery part turned out to be the easy 20%. The hard part was everything downstream: keeping AI-personalized messages from reading as generic once you send 100+ of them, and not letting outreach slide into spam territory. How are you handling deliverability with AI-generated messages? That was the part that surprised me most.
Yeah, discovery was the easy part for me too. Deliverability and “this still sounds like a person” are where most of the energy went.
What I’m doing, roughly:
Treat AI as a draft factory, not a send button. Every message sits in a review queue. Nothing goes out until I approve it. That alone stops the “100 generic emails before you notice” failure mode.
Keep the send path boring. Plain text only, dry-run by default, daily cap + spacing between sends, send windows, suppressions. No HTML templates, no tracking pixels, no “act now / game-changing” sludge. The model also gets length limits and a banned-phrase check so it doesn’t ship opener spam.
Verify before you burn reputation. I don’t trust scraped/guessed addresses as-is — verify the mailbox first, then send. A “looks right” email that bounces is worse than skipping the lead.
Personalization has to be one concrete hook, not “noticed you’re growing.” If the draft can’t point at something specific, it usually gets rewritten or killed in review.
Still learning the hard way on volume though — curious what surprised you most on the spam side. Was it content scoring, bounce rate, or domains getting toasted?
The “who should I talk to next?” framing is interesting because that’s a different problem from automating the outreach itself.
From people using it so far, is the real value turning out to be better prospect selection, or replacing the LinkedIn/Apollo/AI/spreadsheet workflow around it?