The "Firehose" Problem

Imagine being the only person standing between a mountain of customer inquiries and a peaceful day at the office. That was the reality for the customer service lead at Thomas & Piron during peak activity periods.

The phone wouldn't stop ringing, the inbox was overflowing, and frustration was mounting—both for the customers waiting for answers and the team trying to provide them. High churn was becoming a real threat, and burnout was looming.

We needed a solution. Not just a band-aid, but a digital reinforcement.

Contents

• Enter the AI Agent
  • Core Capabilities
  • Code Highlight: The "Human-in-the-Loop" Pattern
• Under the Hood: The Tech Stack
  • The "Brain" (RAG Pipeline)
• The Impact
• Conclusion

Enter the AI Agent

We didn't just build a chatbot; we built a capable digital assistant. This agent wasn't designed to replace the human touch but to handle the heavy lifting so the human team could focus on what really matters.

Core Capabilities

Code Highlight: The "Human-in-the-Loop" Pattern

One of the most critical features was ensuring the AI didn't go rogue. Here is a simplified snippet of how we defined a Semantic Kernel function that requires human approval before finalizing an action.

[KernelFunction, Description("Schedules a meeting with a human agent.")]
public async Task<string> ScheduleMeetingAsync(
    [Description("The email of the customer")] string customerEmail,
    [Description("The preferred date and time")] DateTime preferredTime
)
{
    // 1. Check availability
    if (!await _calendarService.IsAvailableAsync(preferredTime))
    {
        return "Slot not available. Please suggest another time.";
    }
 
    // 2. Create a 'Pending' appointment requiring approval
    var appointmentId = await _crmService.CreatePendingAppointmentAsync(
        customerEmail,
        preferredTime,
        requiresApproval: true
    );
 
    // 3. Notify the human agent
    await _notificationService.SendApprovalRequestAsync(appointmentId);
 
    return $"Meeting request created (ID: {appointmentId}). Waiting for agent approval.";
}

Under the Hood: The Tech Stack

I served as the Full-Stack Developer on this project, handling everything from the React frontend to the .NET backend. Here is the arsenal we used:

The "Brain" (RAG Pipeline)

Building a RAG system that actually works is more than just throwing text into a database. We had to get creative with ingestion and retrieval.

// Example of the retrieval pipeline with reranking
public async Task<List<DocumentChunk>> RetrieveAndRerankAsync(string userQuery)
{
    // 1. Vector Search (Recall)
    var embedding = await _embeddingGenerator.GenerateEmbeddingAsync(userQuery);
    var initialResults = await _qdrantClient.SearchAsync(embedding, limit: 20);
 
    // 2. Reranking (Precision)
    // We use a cross-encoder model to score relevance between query and documents
    var rerankedResults = await _reranker.RerankAsync(userQuery, initialResults);
 
    // 3. Filter and Return Top K
    return rerankedResults
        .Where(r => r.RelevanceScore > 0.75)
        .Take(5)
        .ToList();
}

The Impact

The results spoke for themselves. By deploying this agent, we managed to turn the tide during peak season.

Conclusion

This project wasn't just about using the latest AI buzzwords; it was about solving a real human problem with technology. We took a stressed-out support process and injected it with intelligence, efficiency, and a bit of .NET 9 magic.

Now, the customer service lead can actually take a lunch break. 🥪

Read another project

If you'd like to explore more, here's another project you might enjoy.

AI-Powered Portfolio

4 min

Discover how I implemented a knowledge-based AI assistant for a web dev portfolio.