Deals-machine

Deals Machine is an autonomous AI sales agent that handles outreach, follow-ups, CRM updates, lead scoring, and scheduling so founders spend time selling, not managing sales ops.

Comment

Inspiration

Sales teams don’t lose deals because they cannot sell. They lose deals because operational work overwhelms them. Founders and small revenue teams spend hours writing follow-ups, updating CRMs, scheduling callbacks, and managing pipelines instead of actually talking to customers. We wanted to build an AI system that removes the operational burden from sales while keeping the human relationship at the center. Deals Machine was created to let humans focus on conversation and trust, while autonomous agents handle everything else.

What it does

Deals Machine is an autonomous AI sales agent that manages the sales pipeline from post-call analysis to follow-up execution. It ingests sales call audio through Speechmatics, transcribes and analyzes conversations, creates persistent lead memory, and autonomously decides the next best actions for every deal.

The platform can:

  • Draft personalized follow-up emails
  • Schedule callbacks and meetings
  • Update CRM activity automatically
  • Score and prioritize leads
  • Re-plan outreach when deals stall
  • Read inbound documents like RFPs, contracts, and technical briefs
  • Use multimodal reasoning on screenshots and uploaded materials

Each lead operates through its own agent loop with memory, tools, and decision-making authority, allowing the system to continuously move deals forward without constant human supervision.

How we built it

We built Deals Machine using:

  • Next.js for the frontend experience
  • Supabase for authentication, storage, and database management
  • Vercel for frontend deployment
  • Vultr Ubuntu 24.04 for autonomous agent workers
  • Speechmatics for speech-to-text transcription
  • AI orchestration pipelines for memory synthesis, lead reasoning, and workflow planning

The architecture revolves around autonomous lead agents. Every lead maintains contextual memory from prior conversations, documents, and activities. Before every action, the agent re-reads its memory, evaluates the deal state, and decides what should happen next.

To improve security, we integrated Veea Lobster Trap middleware to harden the system against prompt injection attacks hidden inside inbound emails or uploaded documents.

Challenges we ran into

One of the biggest challenges was balancing autonomy with reliability. Sales communication is sensitive, and fully autonomous actions must remain contextually accurate and human-like.

We also faced challenges with:

  • Maintaining long-term lead memory consistency
  • Preventing hallucinated follow-ups or incorrect assumptions
  • Designing autonomous retry logic for stalled deals
  • Handling prompt injection risks from external content
  • Coordinating asynchronous workflows across multiple lead agents

Building trust in agent-generated outreach required extensive testing and iteration.

Accomplishments that we're proud of

We are proud that Deals Machine functions as a true autonomous workflow system instead of a simple AI assistant. The platform actively reasons about pipeline state, adapts when deals go cold, and continuously operates without requiring manual prompting.

Other accomplishments include:

  • Building a fully multimodal sales intelligence pipeline
  • Implementing autonomous replanning behavior
  • Creating persistent per-lead memory systems
  • Securing the reasoning loop against malicious inbound content
  • Deploying a scalable architecture capable of supporting multiple concurrent lead agents

Most importantly, we built a system that gives founders back time to focus on actual customer relationships.

What we learned

We learned that autonomy in sales is not about replacing humans. It is about eliminating operational friction. The best results came when AI handled repetitive execution while humans focused on trust, persuasion, and strategy.

We also learned:

  • Memory quality is more important than raw model intelligence
  • Reliable workflows matter more than flashy AI outputs
  • Security becomes critical when agents can take external actions
  • Small teams benefit massively from workflow automation when execution quality remains high

What's next for Deals Machine

Next, we want to expand Deals Machine into a fully adaptive revenue operating system.

Planned features include:

  • Real-time call coaching during live meetings
  • Deep CRM integrations with Salesforce and HubSpot
  • Multi-channel outreach across LinkedIn, SMS, and WhatsApp
  • Predictive churn and deal-risk analysis
  • Autonomous pipeline forecasting
  • Team-level coordination between multiple sales agents
  • Fine-tuned vertical-specific sales intelligence models

Our long-term vision is to create an AI-native sales infrastructure where operational sales work disappears entirely, allowing humans to focus only on high-value conversations and relationships.

Built With

Share this project:

Updates