Why We Built Our Own Outreach Pipeline (Not Another SaaS Inbox)

We already ship production AI — WhatsApp on vessels (1,000+ messages a week), Academy intern workflows, SEO agents on the same FastAPI router. Outbound was the awkward cousin: Hunter for emails, something like Instantly for sends, replies in a Gmail tab nobody checked on weekends.

It worked until it didn't. A prospect replied on Thursday. I found it Monday, buried under newsletters. Same person existed in two lists with slightly different titles — we almost emailed them twice. "AI personalization" was {first_name} plus a sentence scraped from LinkedIn About. Not great.

So we built outreach into the stack we already run. Not because cold email is magic — because coordination and deliverability kept losing to tools that don't talk to each other.

The admin panel (real prod, not Figma)

The GIF below is the actual dashboard we use — outreach funnel, sent prospects, unified email inbox. No mock data.

What was breaking

Replies vanishing. Sends go from [email protected]. Replies should update CRM state. They didn't — they sat in Gmail until someone remembered to look.

Duplicate rows. Two discovery runs, same email, different company spellings. Without a hard dedup gate, you send twice and look spammy.

Deliverability as a one-time DNS task. SPF/DKIM/DMARC green on day one means nothing if you blast 40 emails on a fresh domain at 2am UTC. Warmup has to be code, not a spreadsheet reminder.

Empty inbox bodies. Resend's webhook gives you metadata. HTML-only replies showed up in admin with blank bodies — useless when you're trying to decide if someone is warm.

What we built

Piece	What it does	Why it exists
Discovery	Finds ops-heavy companies via DDG + intent APIs	Rows land as pending, not mystery CSVs
Quality gates	Blocks role emails, directory sites, junk domains	Saves reputation before SMTP
Email dedup	One active row per normalized address	Same person, two lists → one send
Draft agent	LLM writes short, question-led mail	We want a reply, not a pitch deck
Send scheduler	Mon–Fri, 9–11am recipient local time, daily cap	Humans actually read mail in the morning
Unified inbox	Reply → Resend → router → Admin Email + status=replied	Warm lead in one panel, Discord ping included

Cadence in prod

Hourly job: draft up to five pending prospects, send up to three — capped by warmup phase. Sunday night discovery. Monday 09:30 UTC still posts a batch to Discord #outreach-agent for a human skim before anything controversial goes out.

outreach_discover: "Sun 22:00 UTC"
outreach_pipeline: "hourly POST /v1/outreach/pipeline"
outreach_legacy_draft: "Mon 09:30 UTC → Discord #outreach-agent"
warmup_caps: [8, 15, 25, 40]   # by lifetime sends on mail.appopoleis.com
send_window: "Mon–Fri 09:00–11:00 recipient local"

At ~76 lifetime sends: zero bounces, SPF/DKIM/DMARC still green. Slow on purpose. The domain is young; we'd rather send 15 good emails than 150 that torch reputation.

When someone hits Reply

Prospect replies to cold mail → Resend receives it → we pull the full MIME (HTML stripped to plain if needed) → match sender email to an outreach row → flip status=replied → thread shows in Admin → Email.

Recent inbox fixes (the unglamorous stuff that actually matters):

1. Spam sender blocklist — mark once, auto-filter the next time
2. Resend receiving API — full body, not empty shells on HTML clients
3. Attachments — PDFs and proposals downloadable from the panel
4. Outreach filter — [email protected] separated from catch-all noise

Pipeline sketch

Same pattern as everything else we ship

Outreach isn't a side project bolted onto the website. It's the same shape as our WhatsApp handler and LLM router: hard rules before the model runs, audit trail in Supabase, one admin surface instead of five logins.

We're not promising "100 meetings a month." We're building the boring rails — dedup, warmup, bounce pause, reply routing — so when someone actually responds, you notice the same day.

If you're running custom AI already and outreach is still a spreadsheet plus three SaaS tabs, talk to us. We'd rather deploy something that survives your inbox than another demo that looks fine until week three.