• 

CRM Heal — Autonomous CRM Cleanup Agent

💡 Inspiration

CRM data decays faster than most teams realize. Contacts leave companies, emails go stale, and job titles change — yet sales and ops teams continue relying on outdated information. This leads to wasted outreach, lost opportunities, and significant inefficiency.

We wanted to answer a simple question:

What if your CRM could fix itself?

CRM Heal was built to demonstrate an autonomous agent that doesn’t just analyze data — it takes real-world action to clean, verify, and improve it.

---

🚀 What it does

CRM Heal is an autonomous AI agent that:

• Detects duplicates and incomplete records
  
• Enriches missing data from the open web
  
• Verifies contacts using real phone calls
  
• Logs every decision for auditability
  
• Persists cleaned data into a database
  

All of this happens without manual intervention after upload.

---

🏗️ How we built it

The system is designed as an agentic pipeline:

🧠 Reasoning Layer

• Rule-based + Akash-hosted models (optional)
• Determines:
  • what data is missing
  • when enrichment is needed
  • when verification should be triggered

⚙️ Execution Layer

• TinyFish → autonomous web enrichment
  
• Vapi → outbound voice calls for verification
  
• Redis → queue, state, and real-time updates
  

🧾 Control Plane

• Guild.ai (CLI + compatible control plane)
  
• Tracks:
  • agent sessions
    
  • decisions
    
  • audit trail
    

🗄️ Persistence

• Ghost (ephemeral Postgres)
  
• Stores cleaned CRM records and verification stats
  

🌐 Frontend

• React + SSE-based live dashboard
  
• Shows:
  • real-time agent activity
    
  • confidence scoring
    
  • verification status
    

---

🤖 Autonomous behavior

Once a CSV is uploaded:

1. The agent analyzes records for issues
   
2. It enriches missing fields from the web
   
3. If confidence is low, it triggers a phone call
   
4. The call verifies real-world information
   
5. Results are written back into the system
   
6. A full audit trail is recorded
   

No manual steps required.

---

📞 Demo highlight

The key moment:

The agent identifies a low-confidence record → initiates a real phone call → verifies the contact → updates the CRM live.

This demonstrates true autonomy beyond the browser.

---

🧩 Challenges we faced

🔗 Webhook reliability

Connecting Vapi callbacks to a local environment required tunneling and fallback strategies. We implemented a polling fallback to ensure the system remains reliable even if webhooks fail.

🧠 Reasoning without OpenAI

We didn’t rely on proprietary APIs. Instead, we built a deterministic reasoning engine with optional Akash-hosted models, ensuring the system works even without external dependencies.

🧾 Guild integration without API

Guild.ai doesn’t expose a runtime API, so we designed a compatible local control plane that mirrors session tracing and audit behavior.

⚡ Real-time UX

Ensuring smooth, live updates required careful use of Redis streams and SSE, with fallback polling for resilience.

---

📚 What we learned

• Autonomous agents need reliable fallbacks, not just ideal flows
  
• Real-world actions (like phone calls) introduce non-determinism
  
• Observability (audit logs) is just as important as intelligence
  
• The best demos are simple flows that feel magical
  

---

🏁 What’s next

• Deploy as a production-ready service
  
• Add continuous monitoring of CRM health
  
• Integrate more real-world signals (LinkedIn, email validation, etc.)
  
• Expand into a fully autonomous data operations platform
  

---

🔥 Final takeaway

CRM Heal shows that:

AI agents shouldn’t just suggest changes —
they should act, verify, and improve data autonomously in the real world.

Built With

• fastapi
• python