Using OSS models to build AI apps with millions of users — Hassan El Mghari

Key Concepts

Open source models, AI app development, user interface (UI) design, rapid prototyping, market validation, AI model integration, viral loops, serverless architecture, AI infrastructure, LLM analytics, idea generation, product launch.

Intros and Overview

Hassan, leading Developer Relations at Together AI, shares his experience building AI applications. He aims to describe his journey, showcase his apps, explain his development process from idea to launch, and offer advice for others. He emphasizes the historic opportunity for building due to lowered barriers and groundbreaking models.

Together AI

Together AI is a platform for open source models, offering an inference API for querying diverse models like Quen 3 (chat), Deepsecar1 (reasoning), and Flux (image). They provide dedicated instances, fine-tuning options, and a GPU cluster product.

The Historic Time for Building

The combination of lowered building barriers using tools like Cursor, Windsurf, Bolt, Vzero and chat apps with groundbreaking AI models released weekly makes it an ideal time for innovation.

App Demos and Examples

Hassan demonstrates several AI apps:

• AI Glasses Picker: An app that helps users pick glasses using the Amazon API based on requirements.
• AI Commit Message Generator: An app generating commit messages from git diff data. It has 40,000+ users and open-source contributions. Example: Theo requested an AI commit message writer on Twitter, leading to the creation of this popular app.
• Text to App Builder: Converts text prompts into functional React code, generating over a million apps from 5 million requests. Example: User input "quiz app about American history" translates to a project plan, then React code.
• Blinkshot: An app that generates images in real-time.
• Napkins: Builds web apps from napkin sketches screenshots (40,000 users).
• Resume to Personal Site: Creates personal websites from resume uploads.
• AI Tutor App: Explains topics like personal finance at an elementary school level.
• OCR Apps: Optical Character Recognition applications.

Some apps have thousands of users, while others have millions, highlighting the numbers game aspect of finding successful ideas.

Tech Stack and Architecture

Architecture:

1. User Input: Typing or image uploads.
2. AI Model Processing: Input sent to an AI model (usually on Together AI).
3. Data Storage: Results (images, text) stored in a database.
4. Output to User: Displaying processed results.

This simple architecture (often with one API call) is crucial for rapid development and idea validation.

Tech Stack:

• Together AI: For AI models.
• Next.js & TypeScript: Full-stack framework.
• Neon: Serverless Postgres hosting.
• Clerk: Authentication.
• Prisma: ORM (Object-Relational Mapping) for database interaction.
• Shadcn UI & Tailwind CSS: Styling.
• S3: Image storage.
• Plausible: Website analytics for unique visitors, location, device information, surprisingly high mobile usage.
• Helicone: LLM analytics for troubleshooting.
• Vercel: Hosting.

Process for Building Apps

1. Ideation: Maintain a running list of ideas, noting interesting concepts or products. Keep a short list of top five ideas. Example: Noticing a need for an app from a tweet, then fulfilling it.
2. Naming: Choose a short, memorable name. Use tools like Domains GPT to check domain availability.
3. Design: Plan app workflow (landing page, input, output). Sketch on paper, use Figma, or prototyping tools.
4. Building: Create the simplest working version, aiming for a single API endpoint.
5. Authentication & Limits: Consider authentication, usage limits, and API key options based on app and AI model costs.
6. Pre-Launch Prep: Create OG images, acquire a domain, add analytics, and write a comprehensive README for open-source code.
7. Launch: Announce on platforms like LinkedIn or X.

Advice for Building Apps (Seven Tips)

1. Simple, Exciting Idea: The concept should be describable in five words (e.g., "Blinkshot generate real-time images"). Avoid overly complex projects initially. Example: Contrast this with a grandiose personal CRM project.
2. Prioritize UI/UX: Spend 80% of effort on UI. Good UI can significantly increase user adoption, even for simple apps like PDF summarizers.
3. Keep it Simple: Limit to one or two API calls for simplicity and speed.
4. Incorporate Latest AI Models: Using new models (like Flux Chanel for Blinkshot) can increase virality. Launching apps quickly after model releases creates opportunities.
5. Launch Early and Iterate: De-risk by launching quickly and iterating based on user feedback. Don't spend months on an unproven concept.
6. Free and Open Source: Encourages learning and sharing.
7. Keep Shipping: Success is a numbers game. Building more apps increases learning, skill, and the ability to identify promising ideas.

Q&A Highlights

• Why not start a company? Hassan enjoys teaching and providing free, open-source resources. Sponsorships help sustain these efforts.
• Common trends in successful/unsuccessful apps? Apps with easy sharing options (viral loops) tend to perform better.
• How to fund compute and model calls? Together AI sponsors compute. Partner with other AI companies (Neon, Clerk) for free credits in exchange for being part of an open source project.
• Challenges with quickly changing AI models? You can easily update and relaunch apps with new models, a key advantage in AI app development. It can be as simple as "a one-line change"

Synthesis/Conclusion

Hassan's approach to building AI apps centers on rapid prototyping, simplicity, open-source principles, and continuous learning. By leveraging the latest AI models, prioritizing user experience, and embracing iteration, developers can create successful AI applications with millions of users. The key takeaways are to focus on simple, easily understandable ideas, prioritize the user interface, and actively seek opportunities to integrate new AI technologies.