Build an AI Customer Service Agent with DeepSeek-R3 and Netlify

What this is, in plain English

This is the story of a business owner who replaced hiring junior staff with a custom AI agent that answers customer questions. The agent has three parts, described in body terms: a “brain” (the DeepSeek AI model, which thinks up the answers), “hands” (a web page where customers type or speak their question and see the reply), and a “voice/edge” layer (small server programs on Netlify that connect everything and handle speech).

Be honest about the level: this is an advanced, build-it-yourself project. You write code, use an API key, and deploy server functions, and the original author shares the design rather than copy-paste code. So this page explains the architecture, what you can use it for, and the realistic order of building it, so you can judge whether to take it on or use a simpler no-code option.

What you can use it for

• A WhatsApp or Instagram FAQ bot. Answer sizing, delivery, and care questions instantly, even while you sleep.
• An internal IT helpdesk. Handle “how do I reset my password” and Wi-Fi questions so staff stop opening tickets.
• Store and service support. Cover your common questions about products, bookings, or policies around the clock.
• After-hours coverage. Keep answering customers when no human is available.
• Any FAQ-heavy support role where most questions repeat and good answers can be written down.

Tools you need

• DeepSeek-R3 (free): the AI model that answers the questions (the “brain”). You call it through its API with an API key.
• Netlify (freemium): hosts your web page and runs small “serverless functions” (snippets of backend code that run on demand) that glue the parts together.
• Web Speech API (free): a built-in browser feature that turns speech into text and text into speech, for voice input and replies.

How it actually works

You build three connected parts and deploy them on Netlify. The realistic shape:

1. Set up the brain. Sign up for DeepSeek, get an API key, and call the model with the customer’s question plus your business details, so it answers about your products, not in general.
2. Build the hands (the web page). A page with JavaScript that captures what the customer types or, using the Web Speech API, what they say, and displays the AI’s reply.
3. Add the voice/edge layer on Netlify. Create serverless functions that take the question, call DeepSeek, and return the answer. The author had to split the work across functions because Netlify’s free functions stop after 10 seconds.
4. Deploy it. Publish the page and functions to Netlify so the agent is live at a web address.
5. Ground it in your real information. Give the AI a curated set of facts about your business (your FAQ, policies, products). This is the single biggest factor in stopping it from making things up (“hallucinating”).
6. Optionally publish audit logs. Keep a public record of interactions for transparency, as the author did.

Words you’ll see, explained

• API / API key: the API is the online service that runs the AI; the key is your secret pass that lets your code use it.
• Serverless function: a small piece of backend code that runs on demand (on Netlify) without you managing a server.
• Edge: code that runs close to the user for speed; here, the layer connecting the page and the AI.
• Web Speech API: a browser feature for turning speech to text and text to speech.
• Hallucination: when an AI states something made up as if it were true; grounding it in your facts reduces this.
• Audit log: a record of what the agent was asked and answered, kept for transparency.

Original source

Based on a Show HN post by axotopia (“Dumped Wix for an AI Edge agent so I never have to hire junior staff”), describing a custom AI agent built to handle customer service for a building-design consultancy.

Notes & variations

• Do you even need to build this? For simple FAQs, a no-code chatbot, or a custom GPT or Gemini Gem loaded with your information, can do the job without any coding. Build this when you want full control and your own deployment.
• Free-tier alternatives: Vercel offers a free tier similar to Netlify for hosting. Other free or cheap models can replace DeepSeek.
• Common pitfall: relying on the Web Speech API without good error handling, especially for non-English speech and accents, which the author found unreliable.
• Tip for better results: define the agent’s knowledge clearly. The more curated, specific information you give it about your business, the more accurate its answers and the fewer hallucinations.