My AI Tech Stack – March 2025

Several people have asked me about my AI workflow, so I thought I’d share my current “AI tech stack” as of March 2025. Given the unprecedented velocity of innovation in this space, consider this a time-sensitive snapshot. Some of these products may not survive long-term, reminiscent of dot-com era casualties like Webvan, Boo.com, or Pets.com.

Below you will find what I’m using, how I’m leveraging each tool, their associated costs, and my unfiltered thoughts on each.

Note: Those wondering, I am paying ~$115/month + additional dev costs in total for all these tools. It sounds insane, but my work quality has gone up, the number of verticals I can work has drastically expanded and my curiosity has never been higher.

Podcast of this Post (powered by NotebookLM): I turned this post into a podcast with AI hosts and everything. Listen here.

_{TL;DR – Summary by GPT4.5 – AI tools are evolving at breakneck speed, and my current AI stack reflects what’s working for me as of March 2025. I’m using a mix of foundational models (OpenAI, Anthropic, xAI, DeepSeek), development tools (Replit, Cursor), and workflow automation (n8n), among others. While I spend about $115/month on AI subscriptions, the productivity gains, expanded capabilities, and sheer learning potential make it a no-brainer. I also see trends emerging, like model routing and agentic AI, that will further shape the landscape.}

Foundational Model Providers

I’m currently using a broader array of models than any point previously, each with unique strengths that merit their place in my workflow. Before diving in, I should note that I subscribe to the premium tier of each service, which grants access to their flagship offerings. The free tiers tend to be fairly comparable across providers, so the differentiation is less pronounced. For those interested in objective comparisons, I recommend following quantitative benchmarks (such as lmarena.ai) to guide your selection based on specific use cases. Please note, there are a number of ways foundational models are being evaluated and no one leaderboard is the end all be all. Anyone interested in this, do some digging to understand how the models are scored, how it can be “gamed”, and what benchmark is most important for your use case.

Anthropic – Claude Sonnet 3.7

Anthropic’s latest model released a few weeks ago excels at generating exceptionally high-quality code. While I’m not a professional software engineer, Claude Sonnet 3.7 enables me to produce code that could convince many actual engineers that I might know what I’m doing. Anthropic also released Claude Code which is an agentic tool that lives in your terminal and is outputting stellar results but is harder to work with (fun fact, its bricked a few computers). Due to the price and how you interface with it (terminal) I am not using this at the moment. Beyond coding, Claude serves as a go-to for introspective conversations—it demonstrates a therapist-like quality with remarkable depth on personal matters.

Use it for: Coding tasks and nuanced, personable conversations
Skip it if: You don’t write code, already subscribe to competitors, or prefer free models (their free 3.5 model is frequently at capacity)
Cost: $20/month –
Standout feature: Claude Code, MCP (Model Context Protocol) integration
My API usage: Not currently utilizing Anthropic’s API for my applications, though I likely will in the future, especially for code-generation-centric projects

xAI – Grok 3

Elon assembled the world’s largest GPU cluster in record time to train Grok 3. While it didn’t initially stand out as essential, I quickly recognized its strengths in leveraging X’s dataset alongside more permissive guardrails (it’s less politically constrained and not as child-safe). For real-world current events, Grok consistently provides the most insightful analysis. Their product team has also excelled at social app integration, with significant potential for further development.

Use it for: Real-time news analysis with greater candor when you value honesty over diplomacy
Skip it if: You prefer more politically correct responses, question X’s dataset quality, dislike Elon, or need reliable voice interactions (its clunky)
Cost: $20/month
Standout feature: News analysis
API usage: Not currently using Grok 3 API, though I may explore the X API soon (but its expensive)

OpenAI – GPT 4o, GPT 4.5, GPT o1, GPT o3, Dall-E 3, Sora

For personal use, I primarily rely on OpenAI’s GPT 4o and o1 models (not impressed with 4.5 yet but exploring it). The company has maintained leadership in the generative AI revolution (which they arguably initiated). I’ve been using their models since GPT 3.5 launched in December 2022, and their continuous improvements—particularly the “memory” feature—have created a platform that understands me in ways I might not even understand myself. The ability for the model to lean into prior discussions, of which I’ve had thousands, makes OpenAI particularly difficult to replace for writing and research tasks, even if technically superior models emerge elsewhere. I’ve grown especially fond of the “Deep Research” feature, particularly within the o1 model. It’s incredibly effective at conducting in-depth explorations on any topic. I also love the voice mode—it feels natural and intuitive, allowing me to have conversations with the AI much like I would with a colleague, friend, or assistant. It’s a fantastic experience.

As an app developer, I use their API most frequently due to its exceptional ease of implementation, strong performance, and support for structured outputs (JSON)—a critical feature for my work. OpenAI also offers the most natural conversational AI, which I use daily while driving to explore ideas, transcribe thoughts, or capture brainstorms. This conversational fluidity remains unmatched by competitors.

Use it for: A versatile solution that performs exceptionally across most use cases
Skip it if: API costs are a primary concern, open-source is essential, you need specialized capabilities, prefer politically right-leaning responses, or need to fine-tune models for specific applications (e.g., I can’t train Dall-E for my children’s book app due to its restricted architecture)
Cost: $20/month (or $200/month for pro tier, which I don’t subscribe to)
Standout feature: Deep Research capabilities with both 4o and o1, delivering impressive outputs
API usage: Currently using 4o and Dall-E

DeepSeek – V3, R1

If someone had predicted that the most capable open-source model of 2025 would emerge from China, I wouldn’t have believed them—yet here we are. Much has already been written about the ragtag team out of Hangzhou who trained a frontier model on inferior hardware “on the side” for a lot less money than the big guys (the real number isn’t really known btw), so I won’t repeat the story in full, but this is a good reminder just how much the AI space is the Wild West. Things can change in a day. While I don’t frequently use DeepSeek models on a day to day basis, they represent perhaps the most significant development for the research and development community due to their open-source nature. This is my go to with respect to self hosted inferencing.

Important notes: while consumer versions are available for mobile devices, be aware that your data flows to a Chinese company, which may raise concerns. However, you can host the open-source model on any inference provider to maintain complete data control at attractive pricing. For enterprise projects requiring top-tier performance with in-house hosting, cost-efficiency, or model fine-tuning capabilities, DeepSeek or DeepSeek variants hosted on Together.AI (or alternative inference providers) offer compelling solutions.

Use it for: Open-source implementation, quality outputs, cost-effectiveness, fine-tuning projects where test-time compute matters
Skip it if: You’re a consumer simply seeking a conversational partner
Cost: No direct subscription; inference costs are remarkably affordable for a reasoning model compared to proprietary alternatives
Standout feature: Test-time compute capabilities at exceptionally competitive pricing
API usage: Not using DeepSeek’s API directly, though I would consider APIs hosting open-source variants of their models
Cool Nature article on the company & story

Hugging Face – Various Models

Unlike the other companies listed, Hugging Face doesn’t pretrain large language models. Instead, it functions as a community platform for developing, maintaining, hosting, and deploying open-source NLP and ML models. Think of it as an “app store” for pretrained models with excellent fine-tuning tools. I’ve only used a handful of their offerings, such as computer vision models for object detection (specifically for bicycle analysis), but this represents arguably the most powerful model ecosystem globally.

Hugging Face resembles an international AI/ML marketplace where you can find solutions for virtually any specialized need from developers worldwide. Their educational resources rival those of Stanford and MIT for learning about various model architectures and fine-tuning methodologies.

Use it for: Specialized use cases beyond what frontier model companies offer (especially computer vision), open-source implementations, fine-tuning pretrained models (their tooling is exceptional), research, or general exploration
Skip it if: The technical aspects seem overwhelming or you simply want a conversational AI
Cost: Pay-per-compute without a fixed monthly fee
Standout feature: Their powerful tooling enables competitive model fine-tuning with minimal investment and moderate technical knowledge
API usage: Occasionally for sandbox environments, not currently in production applications

Honorable Mentions

Google Gemini: I briefly subscribed to Gemini’s premium tier when they pioneered the “Deep Research” function, but OpenAI quickly matched and surpassed their capabilities in terms of output quality (in my opinion). Their other models are adequate but not compelling enough to maintain my subscription. From a developer perspective, I find Google’s APIs frustratingly complex—clearly designed for enterprise-scale applications rather than startups seeking simplicity. Their documentation often requires pausing projects to essentially take a course in their API architecture.

Meta’s Llama (all versions): Meta is currently a bit off the back with respect to their foundational models offering. While Llama was originally the go-to open source LLM for researchers and developers to use for any specific purpose (post training it is still super common), other more powerful, more efficient and less restrictive models have emerged, the most notable being DeepSeek R1 and V3. As a result, Meta’s offerings feel incredibly vanilla and non differentiated. Do not misread, there are going to be a high number of models that still use some piece of Llama in the open source community, but considering the size of Meta, the amount of money they have sunk into generative AI and where they stand in the race, I think its a good bet that Zuck is doing some head scratching. Everyone will continue to keep this company on their radar, but for now, Llama is not something I’m using for any workflow or app.

Alibab’s Qwen: I have honestly used Qwen’s models zero times, but this team is making a bit of a stir the last few weeks so its worth having them on the radar. The progression out of China is undeniable, Alibaba yet another company we have to keep on our radar through the rest of the year.

Perplexity: Though functioning more as a router to other foundation models than a core provider, it deserves mention. I initially enjoyed Perplexity, but as foundation model companies have increasingly integrated web search with model summarization and analysis, the product never achieved stickiness for me. My limited experience with their API for product development has been underwhelming.

Note: Looking ahead, I anticipate “model routing” and “pay-as-you-go” solutions to continue to emerge, such as OpenRouter. IE, whatever you are trying to get a model to do, a decision will get made by the “router” (in this case, openrouter uses notdiamond.ai) as to how to best route your task/question/prompt/request to the best model and you’ll pay “as you go” as opposed to guessing what service to use and having a bunch of monthly subscriptions.

Software Development

This category has impressed me most with its AI-enabled innovations. I anticipate an explosion of custom solutions being built for (or at) companies, a shift away from SaaS, software margin compression and costs broadly drop. I also expect a lot of junk to get put out there.

In any event, while I’ve explored numerous options, two stand out as exceptional:

Replit

Replit transforms prompts into functional software within minutes. Though not flawless, it’s the closest thing to coding magic I’ve encountered since building my first computer three decades ago. What makes Replit extraordinary is its ability to quickly iterate toward working prototypes, sync with GitHub, transition to Cursor, engage professional developers through their bounty program, or deploy to scalable pipelines—all from a single cloud-based platform. The main drawback is potential cost escalation, particularly when the code-generating agent encounters and struggles with bugs. I’ve already written a bit about my experience here, but I must note it has improved significantly since they deployed Claude 3.7 into their flagship product.

Use it for: App development experimentation, learning coding principles, and rapid prototyping when time and budget permit
Skip it if: You lack patience or have no interest in understanding software fundamentals
Cost: $25/month base plus usage (I currently spend approximately $100/month while producing numerous applications)

Cursor

Imagine taking the world’s best IDE, forking it, and integrating superior AI capabilities—the result is a company achieving remarkable success, reportedly surpassing $100M ARR with just 12 employees. What’s particularly striking is that Microsoft, despite possessing all the necessary components (codebase, engineering talent, and AI focus), missed this opportunity. Instead, a Y Combinator-backed startup has redefined the IDE landscape. While their long-term viability remains uncertain, they’re positioned to transform the IDE market similar to how Slack revolutionized workplace communication—demonstrating how subtle differences can dramatically impact productivity tools.

Use it for: Professional software development with AI integration
Skip it if: You’re unfamiliar with IDEs
Cost: Variable; I pay $25/mofth

My Cursor + Replit Workflow

My typical workflow begins in ChatGPT or Claude where I work on a PRD (product requirements document) for whatever I’m aiming to build. This usually requires a lot of back and forth, revising, thinking, head scratching and meandering to get to a place where I have a crisp aim for my MVP. I also usually have two PRDs, one for me, and one that I’ll let the agent work on that is a bit shorter and keeping context windows in mind. From here, I usually go to Replit, where I quickly prototype the PRD using AI prompting. This approach is especially useful for uncovering what I don’t yet know from a product design and management perspective. In a way, the initial phase of prototyping is a bit like an engineer using 3D printing to see for fit, fitment and overall how something looks in the real world. After I get a direction, I’ll work in Replit until the prototype reaches a point where Replit starts to struggle or become inefficient and I will transition to Cursor. To do this, I first create a GitHub repository in Replit and sync my project to it. I then open Cursor and pull the codebase from GitHub directly into Cursor’s IDE using best git practices. Cursor offers significantly more flexibility and power compared to Replit, but it also introduces additional complexity and risk. Mistakes—particularly around basic Git operations—can easily disrupt or damage your codebase, so caution and solid Git knowledge are essential when working in Cursor.

Those wondering, when it comes to app deployment and infrastructure, I’m generally using Replit’s pipes, being everything I’m building is either a proof of concept/MVP or a one off app for a client that will have limited use. That said, I am leaning into AWS for any data warehousing needs, Neon for database (via Replit usually) and Digital Ocean for app hosting if for whatever reason I decide I don’t want to use Replit’s deployment infrastructure.

Side note – Google Colab: Though not AI specific, I use Google’s hosted Jupityer notebook for ‘quick’ data science work, or when I am learning with Hugging Face. Its not something I use every day, or even every month, but its in my stack and a great tool for someone who needs to play with Python, data, and maybe Hugging Face’s transformer library.

Honorable Mentions: Bolt, Loveable, Rork, Windsurf. These merit attention for future potential, but for now, I’m satisfied with the Replit + Cursor combination.

Audio

Those familiar with me know my passion for podcasts, conversations, and voice-based media. Three audio AI products have particularly impressed me over the past 18 months. While I don’t use all of these daily, each shows tremendous promise, and I believe such interactions will become increasingly prevalent.

ElevenLabs

I’ve only used this service occasionally, but like Cursor, they’re experiencing remarkable growth. They provide user-friendly tools for building and training audio agents applicable to various scenarios: customer service, debt collection, AI therapy, or even creating digital doubles for meetings you’d prefer to skip. Their platform offers an intuitive interface, reasonable pricing, and comprehensive capabilities for audio AI development.

Use it for: Any audio AI development or podcast production (this powers Lex Fridman’s real-time translations between world leaders)
Skip it if: You don’t create audio content or build audio applications
Cost: Varies but affordable; I pay $11/month

Google NotebookLM

The “robot podcast overview” I posted atop this blog post is powered by NootbookLM (obviously). This brilliant product converts uploaded documents, PDFs, notes, and other materials into podcast format for learning about subjects of interest. Even more impressively, you can “call in” to your custom podcast and converse with the hosts. While the product is exceptional, its placement under Google’s management suggests it may face a gradual decline without reaching its full potential or market fit. The concept, however, presents an opportunity for entrepreneurs to adapt, refine, and potentially build a viable business.

Image Generation

Beyond my blog, I have only one use case for image generation, which I’ll likely discontinue due to market competition (children’s book generator). During my brief deep dive into various text-to-image models, I gained these insights:

Dall-E (OpenAI)

This OpenAI product delivers moderate results—neither exceptional nor poor. My primary frustrations stemmed from limited control over fine-tuning and constraints when attempting to use it as an illustration generator for my children’s book creator. Costs were relatively high, though this is common across image generation services.

Leonardo

This platform offers quality image generation with a user-friendly interface, solid developer tools, and the most control available without venturing into open-source fine-tuning. Pricing is reasonable, and output quality is strong. I would consider returning to this service for any project centered around image generation.

Workflow Automation

n8n

By far the most powerful no-code workflow automation tool I’ve ever stumbled upon. This is like Zapier, just 100x more powerful. If you have repetitive tasks and would like the robots to help you with them, or have an idea where you need one data source to talk to an AI and then put the output somewhere (such as slack) give n8n a try. I’ve had colleagues give this a go who cannot write one line of code and have incredible (game changing) success.

On My Radar – Technologies of Interest (Not Currently Using)

Manus: X is ablaze with talk of yet another AI company out of China that appears to be outdoing the frontier labs’ best efforts. From what I can tell, Manus is designed to autonomously handle complex, real-world tasks—like building a website, analyzing stocks (e.g., Tesla), planning a Japan trip, or comparing insurance policies—and deliver complete results. One of my favorite X personalities, Deedy Das, suggested it is not hype but very much something to watch out for, which is why I’m putting it on this list and have signed up to receive an invite code (its invite only).

VoyagerAI: While I don’t currently need embedding models or retrieval/reranking capabilities, I appreciate their potential for easily deploying technology to extract insights from unstructured data. The combination of VoyageAI + RAG + AI Model represents significant value creation, likely powering platforms like Replit.

Slidedeck Creation Tools: I am yet to use any of the AI driven deck tools, but see the promise. Contenders include Beautiful.ai, SlideSpeak, Gamma etc – no idea which are good and which suck.

Video Tools: I haven’t yet utilized Sora (OpenAI) or Step Video T2V (open-source video generation). Though I lack immediate need, these tools offer tremendous value for digital advertising—particularly the ability to integrate products into AI-generated video, eliminating a major bottleneck for advertisers: content creation. Several startups are developing in this space.

Music Production Tools: With all these tools, I have no time to play with video or audio production the way I used to. This is unfortunate because I’d love to see what is out there for creating fun songs with AI.

Model Context Protocol: Currently trending on X, this open standard enables AI systems to connect with various data sources and tools, providing a standardized framework for accessing external information and performing actions. Essentially, it unifies APIs, data, AI, and “agentic AI.”

Deeper Hugging Face Exploration: I’ve barely scratched the surface of Hugging Face’s capabilities but recognize its immense potential. Fine-tuning open-source models using their toolset represents a vertical with significant value-add potential for enterprises or AI-focused startups. Importantly, differentiation doesn’t always require post-training, but for specific use cases (like my financial news tool), it can provide substantial advantages.

Video Logging and AutoEdit: As far as I know, this does not exist but I anticipate a team is out there working on a tool that runs a model through your database of video (say, my 10 TB of bike/ski/sled/dirt bike/etc footage) and builds a database around this with high level logging of what the video or image (can work for photo too) is. This would require on prem hardware that probably outstrips what most anyone has (though the new Nvidia PC is prime for this). This will open the door to prompt based video editing, which will be a total unlock for the space.

What did I miss? Anything on your radar I should know about!?