Why Most Voice AI Agents Fail - And How We Built One That Works

2025-06-13

In this post, I show you how we built a production-grade real estate voice assistant that actually handles messy, real-world conversations—interruptions, vague answers, multi-step queries, and more. We break down the full architecture—from voice capture (VAPI) to semantic property search (ChromaDB) to automated scheduling (Google Calendar via n8n)—and explain why we chose each tool. Whether you're a developer, founder, or just curious about real-world AI automation, this is the blueprint for building voice agents that don’t suck.

Agent Architecture and Prompt Engineering for Voice AI

2025-06-30

Learn how we designed a single-agent voice assistant that routes tool calls, shapes responses, and keeps conversations natural — all guided by a carefully crafted system prompt. We’ll break down how the agent communicates with tools like property search and scheduling while maintaining a human-like tone. This post is packed with real code, prompt examples, and architectural decisions that you can adapt to your own AI projects.

Handling Dates in Natural Language Without Hallucination

2025-07-28

Learn how our voice AI assistant reliably interprets fuzzy user phrases like “next Saturday afternoon” without hallucinating dates or times. We’ll walk through how we sanitize natural language, anchor parsing to timezones, and guard against ambiguous results — all using code you can plug into your own AI system. This is real-world date handling, built for production use in voice interfaces.

Semantic Property Matching with ChromaDB

2025-07-14

Discover how we used ChromaDB and OpenAI embeddings to match users with real estate listings based on meaning — not keywords. We walk through the entire flow: loading listings, crafting natural summaries, embedding user preferences, and ranking results for voice-based recommendations.