<h2>The Problem</h2>
<p>Charities face a painful dilemma with phone fundraising. Personal phone calls convert donors well, but call centers are expensive and don't scale. When campaigns spike, organizations either miss opportunities or burn through their budgets on temporary staff.</p>
<p>The deeper problem isn't just cost. It's also consistency. Different callers deliver different experiences. Some are great, some aren't. Tracking what actually works across thousands of calls is nearly impossible.</p>
<h2>What I Built</h2>
<p><a href="https://callhenk.com">Henk</a> is an AI voice platform that enables charities to have natural phone conversations with donors at scale. The AI adapts to what donors say, handles objections gracefully, and logs everything for compliance and analysis.</p>
<h2>The Journey</h2>
<p>I underestimated this project completely. I thought voice synthesis exists, telephony APIs exist, just connect them. What I didn't realize was how different voice interactions are from text.</p>
<h3>Tech Stack</h3>
<table>
<thead>
<tr>
<th>Layer</th>
<th>Technology</th>
<th>Why I Chose It</th>
</tr>
</thead>
<tbody>
<tr>
<td>Frontend</td>
<td><a href="https://nextjs.org/">Next.js 15</a></td>
<td>App Router with server actions</td>
</tr>
<tr>
<td>Database</td>
<td><a href="https://supabase.com/">Supabase</a></td>
<td>PostgreSQL with real time subscriptions</td>
</tr>
<tr>
<td>Voice AI</td>
<td><a href="https://elevenlabs.io/">ElevenLabs</a></td>
<td>Most natural sounding synthesis</td>
</tr>
<tr>
<td>Telephony</td>
<td><a href="https://www.twilio.com/">Twilio</a></td>
<td>Programmable Voice, reliable infrastructure</td>
</tr>
<tr>
<td>Hosting</td>
<td><a href="https://vercel.com/">Vercel</a></td>
<td>Edge functions for low latency</td>
</tr>
</tbody>
</table>
<pre><code>flowchart LR
    subgraph Input["Inbound"]
        P[Phone Call] --> T[Twilio]
        T -->|"audio stream"| SR[Speech&#x3C;br/>Recognition]
    end

    subgraph Processing["AI Processing"]
        SR -->|"text"| AI[GPT-4]
        AI -->|"response"| EL[ElevenLabs]
    end

    subgraph Output["Outbound"]
        EL -->|"synthesized audio"| T2[Twilio]
        T2 --> P2[Phone Call]
    end
</code></pre>
<h2>What Building This Taught Me</h2>
<h3>1. Latency is everything in voice</h3>
<p>In text chat, a 2 second delay is fine. On a phone call, 2 seconds of silence feels broken. I became obsessed with latency. Every millisecond mattered.</p>
<p>I learned to pre buffer, predict likely responses, and stream audio before full sentences complete. The realization: real time isn't a feature, it's a completely different engineering discipline.</p>
<pre><code>gantt
    title Latency Budget (must stay under 2 seconds)
    dateFormat X
    axisFormat %Lms

    section Pipeline
    Speech Recognition    :0, 500
    AI Response           :500, 1300
    Speech Synthesis      :1300, 1700
    Buffer                :1700, 2000
</code></pre>
<p>Resources that helped me optimize:</p>
<ul>
<li><a href="https://elevenlabs.io/docs/api-reference/streaming">ElevenLabs Streaming Guide</a></li>
<li><a href="https://www.twilio.com/docs/voice/media-streams">Twilio Media Streams</a></li>
<li><a href="https://developer.mozilla.org/en-US/docs/Web/API/Web_Audio_API">Web Audio API</a> for client side audio handling</li>
</ul>
<h3>2. Real conversations break your assumptions</h3>
<p>I built the happy path first. User asks question, AI responds. Easy. Then real users happened.</p>
<p>They interrupt mid sentence. They mumble. They say "uh huh" while you're talking. They ask questions you never anticipated.</p>
<p>I learned that conversation design is closer to game design than software engineering. You're designing for unpredictable human behavior, not deterministic flows.</p>
<p>Resources on conversation design:</p>
<ul>
<li><a href="https://developers.google.com/assistant/conversation-design">Google's Conversation Design Guide</a></li>
<li><a href="https://www.voiceflow.com/blog/conversation-design-best-practices">Voiceflow's Best Practices</a></li>
</ul>
<h3>3. Compliance isn't a checkbox, it's architecture</h3>
<p>Fundraising has strict regulations. Call timing rules. Mandatory opt out language. Disclosure requirements.</p>
<p>I originally planned to add compliance "at the end." That was naive. Compliance touches every part of the system: recording consent, timing restrictions, script requirements. Building it in from day one would have been easier than retrofitting.</p>
<h3>4. Twilio's documentation tells half the story</h3>
<p>The <a href="https://www.twilio.com/docs/voice">Twilio docs</a> show the happy path beautifully. What they don't tell you is how WebSocket connections drop under load, how DTMF detection varies by carrier, or how international calls have completely different latency profiles.</p>
<p>I learned more from debugging production issues than from any documentation.</p>
<pre><code>// What the docs show:
twiml.say('Hello!')

// What production taught me:
twiml.say(
  {
    voice: 'Polly.Matthew',
    language: 'en-US',
  },
  'Hello!'
)
// Plus: error handling, fallbacks,
// carrier-specific quirks, timeout handling...
</code></pre>
<h2>The Bigger Realization</h2>
<p>Building Henk changed how I think about AI products. The AI isn't the product. The experience is.</p>
<p>Nobody cares that ElevenLabs generates the voice or that GPT-4 writes the responses. They care that the conversation feels natural and the campaign converts donors. I spent too much time on AI capabilities and not enough on conversation design.</p>
<h2>What I'd Do Differently</h2>
<p>Build a conversation simulator from week one. I tested with real calls, which was slow and expensive. A simulator that could throw edge cases at the system would have caught issues faster and cheaper.</p>
<h2>References</h2>
<ul>
<li><a href="https://elevenlabs.io/docs">ElevenLabs Documentation</a></li>
<li><a href="https://www.twilio.com/docs/voice">Twilio Programmable Voice</a></li>
<li><a href="https://supabase.com/docs/guides/realtime">Supabase Real-time</a></li>
<li><a href="https://developers.google.com/assistant/conversation-design">Google Conversation Design</a></li>
<li><a href="https://www.twilio.com/docs/voice/twiml">TwiML Reference</a></li>
</ul>
<h2>Links</h2>
<ul>
<li><strong>Website:</strong> <a href="https://callhenk.com">callhenk.com</a></li>
</ul>


Projects.

Henk: AI-Powered Voice Fundraising Platform