Build & Sell AI Voice Agents From Scratch (Free 4 Hour Course)

Summary

Shreyas Raj presents a beginner-to-advanced tutorial on building, deploying, and selling AI voice agents without traditional coding. The course uses Dograh AI (also rendered as “Dogra” in the subtitles), n8n for workflow automation, and Cal.com for appointment booking. A central theme is minimizing platform and calling costs while understanding the underlying system well enough to troubleshoot it.

1. How voice agents work

The presenter describes a conventional voice-agent pipeline as three components:

• Speech-to-text (STT): Turns incoming audio into text.
• Large language model (LLM): Interprets the text and generates a response.
• Text-to-speech (TTS): Converts the response back into audio.

Each stage contributes to both cost and latency. He recommends aiming for roughly 500–1,000 ms of response delay, with around 800 ms as a practical sweet spot. Very slow responses can feel unnatural, while extremely fast responses may interrupt callers. Telephone calls can add latency compared with web-based audio because the audio travels through a carrier network.

For conventional multi-model setups, he recommends fast, lower-cost models such as Claude Haiku, Gemini Flash, and GPT-4o mini, and advises against using heavyweight reasoning models for ordinary voice conversations. He also introduces real-time speech-to-speech models, which combine listening, reasoning, and speaking in one model. In his experience, these can reduce latency and improve prompt adherence, though their costs should still be tested.

2. Dograh AI platform and cost considerations

The course tours Dograh Cloud’s main areas, including voice agents, campaigns, models, telephony, tools, files, recordings, API keys, agent runs, billing, and reports. The presenter favors Dograh because it is open source and, in his assessment, can avoid the added platform fees charged by some hosted voice-agent services. He recommends starting with the cloud version and using its included credits for practice before paying for external model APIs.

The tutorial also covers:

• Bringing your own model-provider API keys.
• Adding knowledge files, such as PDFs, while noting that large knowledge bases can increase response latency.
• Creating outbound campaigns from CSV files, mapping phone-number fields, setting concurrency and retry behavior, restricting call schedules, and using a circuit breaker to stop failing campaigns.
• Provider billing practices: some providers round calls up to a full minute, even when the call is very short. Exact billing depends on the provider.

3. Inbound booking agent and tool calls

The main build is an inbound receptionist for a fictional personal-injury law firm. The presenter uses Dograh’s node-based agent builder and explains start-call nodes, greeting options, agent nodes, prompts, tools, and global instructions.

His prompting guidance includes giving the agent a clear role, call purpose, and personality, and enabling interruptions so callers can speak over the agent. He demonstrates using an AI assistant to draft and revise a prompt, while cautioning against unnecessary material that could increase prompt size or complicate tool behavior.

For appointment booking, the tutorial connects:

• Dograh tool calls to an n8n webhook using an HTTP POST request.
• The n8n workflow to Cal.com’s API to create a booking.
• Values extracted during the call—such as name, phone number, email, and preferred appointment time—to the request as parameters.

The presenter explains JSON as a structured way to pass data between systems and shows how API documentation and AI assistants can help identify required fields. He works through common implementation problems, including missing or incorrectly formatted fields, phone numbers without country codes, undefined email values, date and time formatting, and the agent not knowing the current date.

His suggested fixes include using the expected time format, passing the current time into the agent’s context, and validating the API request before testing it on a live call.

He then adds a calendar-availability check. The example retrieves upcoming bookings from Cal.com, converts timestamps to the local time zone, removes irrelevant booking details, aggregates the remaining data, and sends a concise summary back to the voice agent. The agent can then avoid offering already-booked slots. The tutorial emphasizes publishing workflow changes and configuring the webhook to wait for a response before continuing.

4. Connecting a phone number

The presenter demonstrates connecting a provider account to Dograh, buying a number, and assigning it to the inbound agent. He discusses Vobiz and VoiceLink, comparing their number costs, activation requirements, and billing increments. These prices and provider comparisons reflect his own experience and are not universal guarantees.

After configuration, he places a live test call and demonstrates the agent collecting information, checking availability, and booking a consultation.

5. Outbound campaigns and a CRM build

The course first shows a basic outbound campaign: upload a CSV, map its phone-number column, configure concurrency and retries, and start the campaign. The presenter cautions against treating outbound AI as a reliable cold-sales closer. Instead, he suggests using it to follow up with people who have already expressed interest, qualify them, and arrange a conversation with a human.

The larger demonstration is Sun CRM, a solar-business CRM built with an AI coding assistant and Dograh as its voice backend. The presenter describes it as a multi-tenant, SaaS-style application and demonstrates:

• Lead management and a pipeline/kanban board.
• Call history, recordings, and transcripts.
• Forms that can trigger automated follow-up calls.
• Outbound calling and campaigns.
• A voice-agent studio connected to Dograh.
• API, webhook, and MCP integrations, intended to let other systems and AI tools interact with the CRM.
• An admin console for organizations and users, including support impersonation with a visible support-session indicator.

The build also includes deploying services to a Google Cloud virtual machine. The presenter covers VM sizing, SSH access, external IP configuration, and firewall rules. He also describes a free-credit option for experimenting with a hosted n8n instance. The CRM build required substantial troubleshooting, and he says he switched from one coding model to another after encountering repeated problems.

6. Optimizing per-minute cost

The presenter argues that a higher per-minute price does not necessarily mean proportionally better voice quality. He recommends testing the balance between response quality, latency, and cost, and favors real-time Gemini models for some use cases. He estimates that a Gemini Live setup can cost around 3–6 cents per minute, depending on prompt size and usage. This is his estimate; actual costs may vary with model pricing and configuration.

He explains that model charges can involve:

• The system prompt.
• Incoming user audio or its interpreted input.
• The model’s generated response.

His practical recommendation is to keep prompts concise—roughly 3,000–5,000 tokens in his tests—and use a tokenizer to check prompt size. He also warns that testing can consume credits quickly, especially with larger prompts or more expensive models.

7. Selling and finding clients

The final modules focus on selling voice-agent services:

• Build confidence through technical understanding: Understand the system well enough to diagnose client problems and discuss requirements.
• Use niche, problem-specific demonstration videos: Show a concrete solution for a particular business type or pain point rather than focusing on subscriber counts or broad topics.
• Lead with value in outreach: Offer a tailored demo or useful automation before asking for a meeting. The presenter also discusses personalized email copy and the costs and setup involved in larger-scale email outreach.
• Avoid relying on cold calls: He advises against making cold calling the primary approach, particularly for people who may face accent or trust barriers in their target market.
• Qualify prospects and focus on outcomes: Position the service around time saved, missed leads captured, or other business benefits rather than simply selling “AI.”

The course also briefly covers sales-call framing, personal branding, and the presenter’s recommendation to use educational content as a source of inbound prospects.

Guides and tutorials covered

1. Voice-agent architecture, latency, and model selection.
2. Dograh Cloud platform orientation and campaign setup.
3. Inbound receptionist creation and prompt design.
4. n8n webhook workflows and Cal.com booking integration.
5. Calendar availability checks and date/time handling.
6. Telephone-provider setup and inbound call testing.
7. CSV-based outbound campaigns.
8. A solar CRM with calling, forms, pipeline management, APIs/MCP, and admin controls.
9. Cloud VM deployment basics.
10. Per-minute cost and prompt-size optimization.
11. Content marketing and value-first client outreach.

Main speaker and source

• Shreyas Raj — Presenter and instructor throughout the video. He shares his own experience and business opinions alongside the software demonstrations.