Powering AI

1. Lights On

What’s the biggest question we’re stuck asking ourselves when we want to build an AI-based application? Where do I even start?

If you’ve had a tango or two with web-based applications before, you know that the process is fairly straightforward — just settle on a tech stack and get to work.

But with AI applications, and especially agentic AI, progress has outpaced the standardization of tools. In other words, we’re stuck at step 0 of the process.

But fear not! There’s now some method to the madness that we can take advantage of to figure out just what exactly is going on.

2. According to Plan

There’s several attempts at classifying tools in AI-space into neat vertical layers — some even go so far as to introduce horizontal classification in the mix by separating between purpose within a layer.

We’re going to avoid belaboring and focus on the core ideas at play.

So, there’s really four main things to worry about in the AI tech stack

• Data — this is concerned with acquisition, storage, and retrieval of data relevant to your AI application.

• Model — the actual AI component(s) doing the heavy work.

• Control — the section of the application responsible for systematic monitoring and coordination of everything beneath.

• Interface — how you and I will actually get to interact with the application.

Now of course, each layer admits its own set of niches — but for the most part, this is sufficient to start charting the AI application space.

For those looking for a visual summary of the tools available at each layer:

3. Scanning for Data

The lowest layer is all about how you handle data.

This is arguably the most important layer of the AI tech stack, as getting it wrong represents a fundamental misunderstanding of the problem your application is trying to solve.

The model you run is only as good as the data it runs on — there’s no point in me writing an agent for handling customer qualification if I can’t figure out how to represent customer data.

Correctness at the data layer requires having a firm understanding of your scenario and knowing all the assumptions at play.

Are you a big business trying to use an AI agent to oversee some administrative tasks, or are you using an AI agent as a virtual salesman on an eCommerce platform?

To that end, there’s quite a few toys to play with. Not all are equal, and some may have to work in tandem. That said, the following may catch your fancy:

• MongoDB is a well-known classic, and for good reason — it’s a pretty good NoSQL database that’s established a stronghold in the applications realm.

• Databricks is a data-centric platform brought to you by the minds behind the beloved Apache Spark. Promising a fully controlled and scalable data warehousing ecosystem built upon a lakehouse architecture, it’s a strong contender.

• LanceDB is actually a new entry to the scene. An open sourced vector database with AI developers as its first priority, LanceDB promises to be the right tool for your RAG needs.

• Chroma is another data storage contender that steps on the mat with embeddings and multimodal support out of the box.

• Qdrant is a vector database hooked with a semantic search engine, allowing you to attain and consolidate information from purely unstructured data!

And that’s just the tip of the iceberg!

4. Artificial Man

The next layer is all about setting up and facilitating the model powering your AI application. This layer is in fact highly coupled with your data layer.

Of course, models work in various ways — LLMs, for instance, are good for text-based problems such as holding up a conversation or answering questions.

Image generators — as the name suggests — use the given information (e.g. a textual description of a scene, the previous frame of a movie, etc.) to whip up an image that tries capturing the spirit of the request.

Let’s see just what kind of models we’re dealing with, shall we?

• OpenAI — you can’t have any discussion about AI anymore without mentioning them. These are the folks behind ChatGPT and GPT-4o — some of the most powerful LLMs we’ve seen so far. Despite completely contradicting their name, they’re worth a buy for hardcore text work.

• Claude — the brainchild of Anthropic, the Claude LLM models are notable for exhibiting better performance at problems involving logical reasoning and analysis.

• Llama — it’s Meta’s open-source shot at the language model scene, with its most recent update being from December last year.

• Mistral — an open-source language model, but quite different in that the tech under the hood is built on what’s called “mixture of experts”. It’s notable for tackling GPT-3 head on.

• Stable Diffusion — an image model brought to you by Stability AI that can conjure up the stuff of dreams.

So far we’re in good hands. But here’s a bit of a brain tease — what if I need more than one of these models?

Craft Thy Prompts

I know you’re lazy, just like me. Hence, I also know how you write your prompts. Basic and without any sharpness.

So we at Antematter built this nifty tool you can use to sharpen up your prompts to foundation models. Give it:

• A layman prompt

• A persona

• Some constraints

And then let it work it’s magic. If you’re interested in trying out the tool, here’s the link.

5. Orchestral Arrangement

It’s not enough to have a model — you need something to help you monitor your model.

And perhaps most interestingly, you need something to help you coordinate several models.

The control layer is the heart of achieving agentic AI — it’s the leap in naive AI applications and modern ones.

By treating AIs as processes that participate in an ecosystem, you get much richer behavior — this is ripe for business opportunity.

Any task you could think of can be “agentified”.

Want reminders for work based on a voice note? Agent.

Want to receive a specialized report on stock trends based on corporate news? Agent.

Want to create a persona that emulates a salesman for your digital store page? Agent.

Want to tackle agents? We have the tools for you.

• LangChain — one of the finest frameworks to enable AI integration into software applications, LangChain powers chatbots, summarization, and RAG.

• CrewAI — a multi-agent framework that orchestrates autonomous AI agents, letting you treat them as highly specialized work units for large scale goals.

• LangSmith — built by the folks behind LangChain, LangSmith lets you focus on the production aspect of agentic applications by providing a framework for debugging, monitoring, and testing.

• ConfidentAI — these guys are tackling model benchmarking and security by providing robust evaluation metrics and strengthened guardrails so that your application doesn’t veer off-course.

6. I See Seas

Last but not least, we actually want to see our application in action and mess around with it at the user-level.

What good is a food order app if it doesn’t let you order food, after all?

At this layer, multiple things contributing to interaction start melding together — authentication, front-end, etc.

The uniting element is the emphasis on enabling interaction.

What do we have here? Well,

• Composio — this lets you connect your AI to over 250+ applications you might be using on the daily!

• Anon — this is a solid framework for connecting your AI to sites without APIs for you to play with.

• Browserbase — it’s a browser that promises to facilitate your AI and agents in surfing the web.

• Arcade AI — a platform dedicated to giving AI versatility by giving them the ability to use ‘tools’ (authenticated integrations) as actions.

Bonus. Languages

Wait, you’re telling me Python isn’t the only language I can work with? Absolutely!

Although languages may receive varying degrees of support, the bottom line is that if push really comes to shove, you can absolutely manage AI and agentic AI with a wholly different language from the mainstream.

• Lua — a beloved classic still used for videogame scripting, Lua was PyTorch’s original home until it migrated to Python.

• C / C++ — we’re committing a cardinal sin by lumping the two, but if you’re working in something intensive like signals in quantitative analysis, you can’t go wrong with the classics.

• Python — the king of the hill, need I say more?

7. Logging Off

We’ve managed to present some of the tools being used in the AI business and placed them in the appropriate sections of an ‘AI tech stack’.

That said, we’re still some ways off — it’ll take a while until there’s an AI-analogue of good ol’ MERN.

If you are building AI agents, we can possibly partner up. Visit our website or schedule a call.