DEV Community: Grant Singleton

I built and open sourced a ts sdk for batch processing LLM calls across model providers

Grant Singleton — Tue, 18 Feb 2025 19:52:05 +0000

I Built a TypeScript SDK for Batch Processing LLM Calls Across Model Providers

Grant Singleton ・ Feb 17

#opensource #openai #typescript #ai

My workflow for going from idea to MVP

Grant Singleton — Tue, 18 Feb 2025 19:51:31 +0000

⚡ The Fastest Workflow to Go from Idea to MVP

Grant Singleton ・ Feb 18

#buildinpublic #webdev #ai #nextjs

⚡ The Fastest Workflow to Go from Idea to MVP

Grant Singleton — Tue, 18 Feb 2025 19:50:33 +0000

If you're about to build a new product, there's a streamlined workflow that can take you from idea to a functional product as quickly as possible. This is my process of going from idea to MVP:

1. Start with a ChatGPT-Powered PRD

Before you write a single line of code, have a conversation with ChatGPT about your idea. Talk through every aspect of it: what the product does, who it’s for, key features, potential edge cases, and technical considerations. Once you’ve covered everything, ask ChatGPT to generate a Product Requirements Document (PRD) based on your conversation.

A solid PRD will serve as the foundation for everything that follows, ensuring that your team (or even just yourself) stays aligned on what you're building.

2. Generate a UI/UX Design Spec Document

Once you have your PRD, start a new conversation with ChatGPT and ask it to generate a UI/UX Design Spec Document. This will outline the structure, layout, and overall experience of your product. It’s important to ensure that the design aligns with the functionality outlined in the PRD.

At this stage, consider asking ChatGPT for best practices around usability, accessibility, and user flows to ensure a great experience.

3. Translate the Design into Figma with shadcn/ui

Now, take your design spec document and build out your product in Figma. Use shadndesign Figma UI Kit, which are pre-built UI components that speed up design work. These blocks provide a strong foundation for an aesthetically pleasing and functional UI.

4. Convert Your Figma Design into Code with the shadndesign Figma-to-v0 Plugin

Instead of manually coding everything from scratch, leverage the shadndesign Figma-to-v0 plugin to generate code directly from your Figma designs. This drastically reduces development time and ensures that your implementation stays true to the original design.

5. Organize Your Project in Cursor

With your UI components ready, move everything into Cursor, an AI-powered code editor (duh). Place your PRD and design requirements documents into a docs folder within your project. This setup allows Cursor’s AI to reference them as you build, keeping development aligned with the original vision.

6. Build Step by Step with AI Assistance

Now, go step by step, asking Cursor to help you build each component and page. Since Cursor can reference your PRD and design docs, it will generate accurate, context-aware code. As you develop, make sure to validate each step, refining the code as necessary.

A helpful tip: Use an ORM like Drizzle so that your database schema is defined in code. This ensures that Cursor has access to your database structure, making it easier for it to generate database-related logic without inconsistencies.

This is the process I used to build Callio and Filtyr. Filtyr took just one week... 🤯

What do you think of this workflow? Do you have any alternative approaches that work better for you? Let me know in the comments!

I Built a TypeScript SDK for Batch Processing LLM Calls Across Model Providers

Grant Singleton — Mon, 17 Feb 2025 21:29:16 +0000

Inspired by Vercel’s AI SDK (But for Batch Processing)

The Vercel AI SDK makes switching models really easy. Just swap the model name, and everything else stays the same. However, the AI SDK doesn't support batching, so I built one for that.

batch-ai gives you a single SDK that works across providers so you can focus on your app, not writing code for different batch APIs.

Why Batch API Calls Matter (Hint: They’re 50% Cheaper)

If you’re processing a high volume of AI requests and don’t need real-time responses, batch APIs can cut your costs in half.

For example, at Filtyr (my AI-powered content moderation SaaS), we process thousands of moderation events daily. By using OpenAI’s and Anthropic’s batch APIs instead of real-time calls, we save 50% on API costs while handling the same workload.

If your use case involves large-scale AI processing, sentiment analysis, classification, content moderation, or research, you should consider using batch APIs.

How to Use `batch-ai`

Here’s how batch-ai simplifies batch processing while letting you switch between providers effortlessly:

import { z } from 'zod';
import { openai, createObjectBatch, getObjectBatch } from 'batch-ai';

// Define output schema using Zod
const responseSchema = z.object({
  sentiment: z.enum(['positive', 'negative', 'neutral']),
  confidence: z.number().min(0).max(1),
});

// Initialize OpenAI model
const model = openai('gpt-4o');

// Batch requests
const requests = [
  { customId: 'review-1', input: 'I absolutely love this product! Best purchase ever.' },
  { customId: 'review-2', input: 'This is terrible, would not recommend.' },
];

// Create batch
const { batchId } = await createObjectBatch({
  model,
  requests,
  outputSchema: responseSchema,
});

// Retrieve batch results at some later point
const { batch, results } = await getObjectBatch({
  model,
  batchId,
});

if (batch.status === 'completed' && results) {
  console.log('Results:', results);
}

Want to switch to Anthropic? Just replace:

const model = anthropic('claude-3-5-sonnet-20241022');

That’s it. No need to rewrite anything else.

Who Should Use `batch-ai`?

If you’re dealing with high-volume AI processing, this SDK can help. Ideal users include:

AI Moderation Platforms (like Filtyr) processing thousands of content moderation events daily.
Marketing Teams analyzing customer sentiment at scale.
Enterprises running classification, summarization, or AI-driven automation.
Researchers working with massive datasets who need structured AI output efficiently.

If you process a lot of AI requests and want to cut costs while simplifying API interactions, batch processing is the way to go.

Future Plans & How You Can Get Involved

I built batch-ai to solve my own batch processing headaches, but there’s more to come:

More provider support: Google Gemini and xAI Grok are next on the roadmap.
Expanding batch capabilities: Adding generateTextBatch for text-based responses.
Better error handling & retries: Making batch requests more robust.

I’d love your feedback! If you have feature ideas or run into issues, open an issue on GitHub.

What’s your experience with batch AI processing? Have you used batch APIs before? Let’s discuss in the comments!

DEV Community: Grant Singleton

I built and open sourced a ts sdk for batch processing LLM calls across model providers

I Built a TypeScript SDK for Batch Processing LLM Calls Across Model Providers

Grant Singleton ・ Feb 17

My workflow for going from idea to MVP

⚡ The Fastest Workflow to Go from Idea to MVP

Grant Singleton ・ Feb 18

⚡ The Fastest Workflow to Go from Idea to MVP

1. Start with a ChatGPT-Powered PRD

2. Generate a UI/UX Design Spec Document

3. Translate the Design into Figma with shadcn/ui

4. Convert Your Figma Design into Code with the shadndesign Figma-to-v0 Plugin

5. Organize Your Project in Cursor

6. Build Step by Step with AI Assistance

I Built a TypeScript SDK for Batch Processing LLM Calls Across Model Providers

Inspired by Vercel’s AI SDK (But for Batch Processing)

Why Batch API Calls Matter (Hint: They’re 50% Cheaper)

How to Use batch-ai

Who Should Use batch-ai?

Future Plans & How You Can Get Involved

How to Use `batch-ai`

Who Should Use `batch-ai`?