Vultr covered the Milan AI Week hackathon and put Deals Machine among the standout winners. What makes their piece worth reading is that it goes past the pitch and into the build: how a human-driven, AI-assisted cold-calling cockpit is actually wired together.
The AI recommends; the rep decides
The framing they used is the one we lead with. Deals Machine sources leads, enriches prospect data, drafts personalized outreach, and coaches live, but every recommendation and every draft goes to the rep for review. The AI never acts on its own, and the human owns every customer-facing decision.
What's under the hood
The article walks through the stack, which is a fair snapshot of how the platform runs today:
- check_circleA Next.js cockpit on Vercel, with hardened Fastify workers on Vultr Ubuntu 24.04.
- check_circleSpeechmatics for real-time transcription and speaker diarization.
- check_circleAnthropic Claude for outreach playbooks, live coaching, and safety classification.
- check_circleTwilio connecting calls into the platform; Supabase storing prospect context, call history, and the evolving memory.
- check_circleA self-improving memory layer that learns from outcomes to sharpen the next call.
- check_circleVeea's Lobster Trap middleware, which screens third-party content for prompt injection before it ever reaches a model.
Every call feeds the memory, so the next call starts ahead of the last one.
That last point is the part we obsess over: the memory layer is what turns a pile of calls into a playbook that rewrites itself. Vultr's piece is a clear outside look at how the pieces fit together.
Read the full write-up on the Vultr blogopen_in_new