El Mghari builds AI apps for millions using a four-step framework

Building AI applications that scale to millions of users requires a shift from complex monolithic thinking to a modular, lean architectural approach. Hassan El Mghari, Lead Developer Relations at Together AI, demonstrates that the most successful apps—like LlamaCoder or Blinkshot—rely on a streamlined four-step flow. The user provides input, the system makes a single targeted API call to an open-source model, the result is stored in a serverless database, and the output is served immediately. This simplicity allows developers to prioritize user experience and speed over complex backend orchestration.

The modern full-stack AI tech stack

To replicate this success, developers must leverage tools that minimize infrastructure friction. The recommended stack focuses on TypeScript-first libraries and serverless scaling:

• AI Inference: Together AI for querying open-source models like Llama 3 or DeepSeek.
• Framework: Next.js with TypeScript for a unified full-stack environment.
• Database: Neon for serverless Postgres and Prisma as the ORM.
• Authentication: Clerk for rapid user management setup.
• UI/UX: Tailwind CSS and Shadcn UI to ensure professional design without custom CSS bloat.
• Observability: Helicone for LLM-specific analytics and Plausible for web traffic.

Code walkthrough for rapid prototyping

The goal is to maintain a single API endpoint. In a typical Next.js Route Handler, you can trigger a model inference and return the stream directly to the client. This reduces latency and improves the perceived performance that users expect from modern AI tools.

import { Together } from "together-ai";

const together = new Together({ apiKey: process.env.TOGETHER_API_KEY });

export async function POST(req: Request) {
  const { prompt } = await req.json();

  // Step 2: Single API call to a high-performance OSS model
  const response = await together.chat.completions.create({
    model: "meta-llama/Llama-3-70b-chat-hf",
    messages: [{ role: "user", content: prompt }],
    stream: true,
  });

  // Step 4: Stream response back to the UI for instant feedback
  return new Response(response.toReadableStream());
}

Seven rules for AI virality

Hassan El Mghari emphasizes that 80% of development time should be spent on the UI rather than the model itself. A beautiful, intuitive interface makes simple functions, such as PDF summarization, feel like premium products. Developers should launch within two days of a new model release to capture the "new tech" trend. By keeping apps free and open-source, builders create a viral loop where users share the output and other developers contribute to the code, effectively outsourcing marketing through community engagement. Success in this field is a numbers game; shipping one app a month is the baseline for finding a hit that resonates with millions.