🧠 Agentic Marketplace: AI-Powered Trust for Digital Deals

🚀 Inspiration

Modern marketplaces make it easy to connect buyers and suppliers—but trust remains a major gap. Deals often rely on fragmented communication, unverifiable updates, and weak enforcement of contracts. This leads to disputes, fraud, and uncertainty.

We were inspired to solve a simple but powerful problem:

What if every deal had built-in intelligence to verify actions, track progress, and enforce trust automatically?

That idea led us to build an AI-powered marketplace pipeline where contracts aren’t just static documents—they’re actively analyzed, monitored, and enforced.

---

🛠️ What We Built

We built a 5-stage deal pipeline that ensures a transaction moves safely from start to finish:

1. Make Contract (AI Risk Scanner)
   Suppliers upload or generate contracts. AI analyzes clauses and flags risks (e.g., liability caps, indemnification imbalance).

2. Agreement Loop (Buyer ↔ Supplier)
   Buyers can accept or request changes until both parties agree.

3. Deal Execution (Evidence Tracking)
   Both parties upload real-time updates (text, images) as proof of progress.

4. AI Analysis
   AI evaluates the full deal:

   • Contract compliance
     
   • Timeline adherence
     
   • Evidence consistency
     

It produces:

• Scores for both parties
  
• Risk findings
  
• A final verdict
  

1. Dispute Window (Escrow Protection)
   Parties can dispute within 24 hours before escrow is released.

---

⚙️ How We Built It

🧩 Tech Stack

• Frontend: Next.js (App Router), React, Tailwind CSS
  
• Backend: Next.js API Routes
  
• Database & Storage: Supabase
  
• AI Model: Gemma 4 (via Google Generative AI API)
  
• File Parsing: pdftotext for robust PDF extraction
  

---

🧠 AI Pipeline

We built a structured AI pipeline instead of relying on raw responses:

1. Extract contract text from PDFs
   
2. Construct a strict prompt enforcing JSON output
   
3. Parse AI output using:
   • Marker-based extraction (FINAL_JSON:)
   • Fallback scanning for valid JSON blocks
     
4. Validate schema before sending to frontend
   

---

📊 Scoring Logic

AI evaluates performance based on:

• Delivery time vs contract:

$$ \text{timeliness score} = \frac{\text{expected days}}{\text{actual days}} $$

• Evidence consistency:

$$ \text{confidence} \propto \text{number of verifiable updates} $$

• Contract compliance:

$$ \text{score} = f(\text{violations}, \text{risk clauses}) $$

---

⚔️ Challenges We Faced

1. Unreliable AI Output

LLMs often returned:

• Multiple JSON objects
  
• Placeholder data ("...", {...})
  
• Partial or truncated responses
  

👉 Solution:
We built a robust parser that:

• Anchors output using FINAL_JSON
• Scans for multiple JSON blocks
• Validates structure before accepting results
  

---

2. PDF Parsing in Node Environments

Libraries like pdfjs and pdf-parse caused:

• ESM import issues
  
• Missing browser APIs (e.g., DOMMatrix)
  

👉 Solution:
We switched to a system-level approach using pdftotext, which proved far more stable.

---

3. State Synchronization Across Pipeline

Managing deal stages (contract → execution → AI → disputes) required careful routing and state consistency.

👉 Solution:
We structured the app around dynamic routes and a shared deal ID, ensuring each stage reflects real-time state.

---

4. AI Determinism vs Flexibility

We needed:

• Consistent outputs (for trust)
  
• But also meaningful analysis
  

👉 Solution:
We tuned the model with low temperature and strict schema enforcement.

---

📚 What We Learned

• AI is powerful—but only when constrained properly
• Parsing AI output is as important as generating it
• Trust systems require both UX and backend rigor
• Real-world reliability > perfect AI responses

---

🌍 What’s Next

• Highlight risky clauses directly in contracts
  
• Improve dispute resolution with AI arbitration
  
• Introduce trust scores across marketplace users
  
• Scale to real-world B2B marketplaces
  

---

💡 Final Thought

We didn’t just build a marketplace—we built a system where trust is enforced, not assumed.

---

Built With

• gemini
• gemma-4
• next.js
• react
• supabase