Before Outvid, running a personalized video campaign meant recording every single video by hand. Two hundred prospects meant two hundred takes, a lost week, and a delivery that got worse with every one. The ceiling was never creative. It was operational. The tools that promised to fix it all broke in the same place. They could generate a video. None of them could send one. No connected sending accounts, no rate limits, no reply detection, no sequence that knew a prospect had already answered. The hard half of outreach was left as an exercise for the customer. So I built the sending engine first and the video engine second. Outvid clones your voice, writes each opening line off live research on the prospect, and lip-syncs it onto footage you record once. Then it sends across email, LinkedIn, WhatsApp and Instagram through a single dispatch layer with per-account and per-domain caps, timezone-aware send windows, reply detection that halts a sequence mid-flight, and bounce-spike auto-pause that shuts a campaign down before your sender reputation goes with it. The architecture assumes you cannot afford mistakes. Workspace-level data isolation enforced on every query. Plan entitlements enforced server-side, not hidden in the interface. Signed, idempotent webhook ingestion. Two interchangeable video providers behind one interface, so no single vendor outage takes your campaign down. Ninety-six data models, five hundred and seventy-two backend operations, typed end to end, with around 160,000 lines of tests underneath.