This whole article was incepted because we decided to do one of the worst things you can do to yourself in business: migrate your entire CRM. If you’ve ever done a CRM migration before, you just winced — because you know exactly what this means. You export everything. You write scripts to transform the data. You map fields, test mappings, find that half the mappings are wrong, fix them, test again. Edge cases everywhere. You do this for weeks. Maybe months. At the end you still find broken records. It’s one of those projects where you clear the calendar and mentally prepare. We had good reasons to do it, but needless to say I wasn’t thrilled about the weeks of cost I was about to incur.

I ended up doing the entire migration in a day. By myself. (and Claude)

Keep in mind, I spent thousands having someone come in and build our HubSpot the first time. It took months. We had hundreds of thousands of records. I just did the equivalent by going back and forth with Claude for 4 hours to make a plan, giving it 2 API keys, and letting it run overnight. It pulled our HubSpot data, reasoned through the Attio schema, mapped the fields, handled transformations, and just ran. When there were ambiguous cases, it asked me. I made judgment calls. The agent executed. Done.

Wednesday afternoon to Thursday evening and our team was fully in the new system. One of the most critical parts of our company, handled with basically zero issues.

That’s what really made me take a step back. My team and I at PressW have spent the last three years pushing the frontier of what’s possible with agents and LLMs, but even I didn’t expect this.

But here’s the thing that made me stop and think: I didn’t really do the migration. I just… directed it. Made a few decisions. The actual work — the data transformation, the API calls, the edge case handling — happened underneath me.

This felt different from other AI tooling I’ve used. It felt less like “AI assisted my work” and more like “I played a video game where the objective was ‘migrate CRM’ and I won.”

That’s a weird way to put it. Let me try to unpack what I mean.

The Software Layer

Here’s a mental model I’ve been developing.

For the past few decades, we’ve been building software that sits between humans and data. Salesforce is a layer between you and your customer data. Jira is a layer between you and your project data. Notion is a layer between you and your documents.

These layers do two things:

1. They provide an interface (buttons, forms, views)
2. They encode workflows (how you’re supposed to do things)

Both of these were necessary because humans can’t interact with raw data. Of course, “raw data” is itself an abstraction — there are layers all the way down to the transistor. The question is which layers still earn their complexity cost. We needed abstraction.

But here’s what I’ve realized: these layers are also constraints. Every Salesforce implementation is basically a negotiation between how your business works and how Salesforce thinks businesses should work. You’re always compromising.

The entire SaaS economy is built on encoding workflows into products. Thousands of companies, billions of dollars, all building layers that constrain how humans interact with data.

What if that constraint is no longer necessary?

Why Code Matters

Okay, so Claude Code (and similar tools) can write and execute code in real-time. This seems like a coding assistant. It’s not. Or rather, it’s much more than that.

Here’s the key insight: code is a universal interface.

Think about it. Every API can be called with code. Every database can be queried with code. Every SaaS product (ultimately) is just code that calls APIs and queries databases. When you give an AI the ability to write and execute arbitrary code, you’ve given it the ability to interact with any digital system that exists.

Previous AI tools operated in constrained domains. A CRM AI works within the CRM. A writing AI works within documents. But an agent that can write code operates at the infrastructure layer. It can reach into Salesforce, pull data, transform it, push it to Notion, update a Linear ticket, and send a Slack message — all in one operation.

This is different in kind, not just degree. You’ve basically built an agent that can do anything a programmer can do, at the speed of an API call.

But (and this is important) code alone isn’t enough. You also need something to tell the agent what to do. This is where it gets interesting.

Skills (or: Workflows Are Just Text Files Now)

So here’s the thing that really changed my mental model.

A “skill” is basically a markdown file that describes how to do something. Decision trees, edge cases, business rules, institutional knowledge — all encoded in a document.

When I first heard this I thought: that can’t work. Complex workflows need UIs, validation, state management. You can’t just describe a workflow in prose and have an AI execute it.

Except… you can. The models are good enough now. Better yet, you don’t even have to write the skill from scratch. The agent can interview you — asking the right questions to extract the workflow from your head. It manages up.

I can write a markdown file that says “here’s how to scope a project: first pull the meeting transcript, extract these signals, check our rate card, look at current team allocation, draft a statement of work with these sections…” and if the agent has access to the right systems, it just does it.

No UI. No six-month implementation. A text file.

This has some non-obvious implications:

Skills are shareable. I write a skill, I share it with my team, now everyone executes that workflow the same way. Institutional knowledge goes from “trapped in someone’s head” to “in the system.”

Skills compose. A proposal-writing skill can call a client-research skill, which calls a transcript-analysis skill. You build complex workflows by composing simple building blocks. This is how learning has always worked — building blocks that stack. Software engineering formalized it, but AI is making it universal.

Skills can be self-improving. The agent can reflect on its own executions, notice where skills fell short, and propose improvements. Or write new skills when it notices repeatable patterns.

That last one is wild if you think about it. The system can literally improve its own capabilities over time based on usage. It’s not static. It learns.

I’ve co-opted the video game terminology and started calling this the “Skills Tree”: workflows as text files, shareable across the org, composable into complex operations, self-improving through reflection.

What We Built

Okay so, this isn’t an article just postulating, we actually built this. It’s called PressW Atlas. It’s been running our company for a few months.

Architecture is simple: Claude Code as the execution engine, MCP connectors to all our systems (Attio, Linear, Fireflies, Gmail, GitHub, Notion, PandaDoc, etc), plus a library of skill files, plus a reflection loop.

When you talk to Atlas, you’re talking to all your systems at once, unified by an intelligence that can synthesize across them.

Some examples of what this looks like:

Proposal generation. Sales call ends, transcript hits Fireflies. Atlas extracts requirements, pulls comparable past projects, checks current team capacity, drafts a full SOW. Our solutions engineer edits instead of creates. Maybe 90% of the work is automated.

Opportunity detection. Atlas monitors meeting transcripts for expansion signals (“we wish we could also…”), watches project timelines, tracks satisfaction indicators. When patterns align, it surfaces opportunities we didn’t ask for. “Client X mentioned Y problem, their current project ends in 3 weeks, here’s a draft proposal.” This kind of cross-system pattern matching was impossible when humans were the integration layer.

Skill creation. I’m explaining a process to someone — say, how we offboard clients. Mid-explanation I realize this should be a skill. I tell Claude to write a skill based on what we discussed. It does. Now that workflow is encoded and anyone can execute it.

The compounding thing is real. When someone on the team finds a better way to do something, they update the skill, and now every agent uses the improved version. Knowledge accumulates instead of staying trapped in individual heads.

What Matters Now (and What Doesn’t)

After running on this for a while, I have some intuitions about what actually matters:

Workflow logic doesn’t matter. This sounds crazy given the entire SaaS industry is built on encoding workflow logic. But if you can just describe a workflow in a text file and the agent executes it… what’s the point of the software? The models are good enough. The constraint was never intelligence.

Data matters a lot. Agents are only as good as their context. If you don’t have meeting transcripts, email threads, project history — the agent can’t synthesize across them. And it’s not just having the data — it’s making it accessible. The balance between what’s pre-structured and what’s synthesized on demand, between what’s captured automatically and what requires deliberate collection. This is where operational advantage will live. The winning companies will be the ones with the richest context layer, not the fanciest models.

Interface is an unsolved problem. If agents do execution, what do humans interact with? Chat feels transitional. Probably something like dashboards showing decisions needed, progress on objectives, opportunities surfaced. We’re still figuring this out. Everyone is.

Security is hard. Giving agents broad access to your systems raises obvious questions. What can they see? What can they do? Audit trails? The security model for agentic systems doesn’t exist yet.

Implications

If this is directionally correct, some things follow:

SaaS is in trouble. The moat of “we encoded this workflow” disappears. If workflows are just text files, why do I need your product? SaaS companies will need to move value elsewhere — to data, to network effects, to being the context layer. Many won’t make it. Of course, the efficient outcome rarely happens cleanly. There are trillions of dollars of market cap, millions of jobs, and entire career paths built on the current stack. The transition will be messy and politically charged — but that doesn’t change the direction.

The gap between adopters and non-adopters will be huge. This isn’t 10% efficiency. It’s a structural change in how work happens. Companies on a Skills Architecture will operate at a different tempo.

What skills matter changes. “Knowing the CRM” becomes worthless. “Knowing what to ask for” becomes valuable. Judgment appreciates, execution depreciates.

New organizational structures become possible. What happens when coordination cost approaches zero? Smaller teams, wider impact. Institutional knowledge that doesn’t leave when people do.

Hard Problems

To be clear, there’s a lot we haven’t figured out:

• How do you manage dozens/hundreds of agents? What’s the UI for that?
• Skills accumulate fast (we have ~100 already). How do you organize/version/deprecate them?
• How do you encode “ask a human” without enumerating every edge case?
• Trust calibration: people verify everything now. That doesn’t scale. And if history is any guide — finance, social media, crypto — we’ll over-trust these systems before we learn their limits. The question is whether we can fail small and learn fast, rather than fail catastrophically.
• Knowledge decay: skills get stale. How do you detect when they’re outdated?

And many more…

These are real problems. I don’t have clean answers. But they’re good problems — the kind you have when something is working and are exciting to challenge.

The Punchline

Here’s the thing about video games.

Games figured out how to give humans control over complex systems without making them manage the complexity. You don’t need to understand the physics engine to play a racing game. The game surfaces what matters — decisions, trade-offs, objectives — and handles everything else underneath.

Enterprise software did the opposite. It exposed the complexity. It made you learn the data model, click through the workflows, manage the state.

The layer between humans and data was necessary for decades because we didn’t have anything else. Now we do.

Agents can be the game engine. They handle the complexity — the data transformations, the API calls, the workflow execution — while humans focus on decisions, judgment, direction.

Work, at its best, has always had this quality — the satisfaction of mastering something hard. Games just figured out how to bottle it. What’s changing is that enterprise software no longer needs to get in the way of that feeling. Not gamification (points, badges, whatever). The actual feeling of fluid control over a complex system.

We’re very early. The tools are rough. But the trajectory seems clear.

The layer is disappearing. The question is whether you’ve noticed — and what you’re going to build on the other side.

Note on PressW

My team and I at PressW have been building production AI systems for companies for over three years now, pretty much started a month after the release of ChatGPT. I’ve spent close to a decade in the AI space, watching the rise from BERT models to vectors to the great LLM wave and now into the agentic future. For the last few years, we’ve obsessed over agent architectures, data, prompting, context engineering, all of it — trying to understand what actually drives value and attempting to glimpse at what the future might look like.

If you’re exploring this same frontier and trying to understand how it might look in your business / fund, I’d love to chat. Shoot me a note: tarun@pressw.ai

2023 has been a groundbreaking year for Artificial Intelligence, arguably the most transformative in human history. This last year no doubt stood out as the Year of the Large Language Model (LLM), fueled by the introduction of ChatGPT in November 2022. This year witnessed an explosion in LLMs and related products, but it wasn’t just LLMs that caught our attention. The advancements in multimodal models, integrations, development tools, and the unfolding drama in regulation and corporate landscapes we found equally captivating.

My name is Tarun Thummala, and I’m the CEO at PressW, an AI agency that now specializes in building custom LLM solutions for businesses. We were founded in January 2023 and since then we’ve had the privilege of working on over 30 projects across 20 clients in 4 industries. Myself and my other two co-founders came with a background in traditional ML and AI, so it’s been fascinating tying our experience with the new age of AI.

We wanted to put this article together to encapsulate the remarkable advancements in AI over the past year and reflect on how they have reshaped our world. In particular we wanted to look at how businesses and the world overall is changing, and how these advancements are fueling this change. Don’t worry, we put this article together to be useful to everyone, regardless of familiarity with AI or LLMs. I will serve as your guide, so without further ado, let’s dive in!

LLM Progress

The Evolution of LLMs

I believe it makes sense to dive straight into how LLMs evolved over the last year. If I had to describe the rate of LLM growth over the last year, I’d have to say: frightening.

Of course the juggernaut, behemoth, raid boss of the space is most definitely OpenAI, led by Sam Altman and his team of brilliant disrupters. For everything that follows, remember, ChatGPT came out in November 2022.

Since its introduction we’ve seen explosive growth in the scene. Within one year, we’ve already experienced the release of GPT-4 and GPT-4 Turbo by OpenAI, Claude by Anthropic, and open-source models like Meta’s Llama.

These newer models are able to process larger amounts of information, answer questions more accurately, and even reason about complex problems better. As evidence by this: GPT-3.5 (ChatGPT) was reported to have an IQ of about 85. GPT-4 on the other hand has recently scored with an IQ of 155. Within one year. We’re pretty sure most people aren’t really grasping how much better these models are getting and how quickly. I’ll sprinkle another mind-bender at you. GPT-4 was created 2 years ago! BEFORE ChatGPT even came out. It’s been two years since OpenAI created GPT-4, now just imagine how powerful their next one will be. We may actually be far closer to superintelligence or AGI than we ever thought.

Multi-Modality

Vision and Sound Take Center Stage

LLM stand for large language model. So it makes sense that most of the applications have focused around written text as data inputs and outputs. But in 2023 we experienced yet another massive jump forward in AI capabilities. Introducing Multi-Modal Models. These are LLMs that are able to accept not only text inputs, but also images, videos, sound, etc. that they can use to factor into their answers.

Check out the prompt belowfor the above image (all autogenerated) from the original input image. Look at how detailed these models are getting at noticing details from pixels and converting it to human language!

{
  "background": "A serene white sand beach with gentle turquoise waves lapping at the shore. Palm trees sway in a light breeze at the edge of the beach, with a vibrant blue sky dotted with wispy clouds overhead. In the distance, a wooden lifeguard stand overlooks the scene, giving a sense of scale and human presence without being obtrusive.",
  "outfit": "A fitted silver metallic suit with a subtle shimmer, featuring a two-button jacket, slim-fit trousers, and a crisp white dress shirt underneath.",
  "user_appearance": "The user has light skin, dark brown hair that is slightly tousled, and a hint of facial hair on his chin and mustache area. He has rectangular glasses with thin frames, and his eyes are partly smiling, suggesting a friendly demeanor. The user's nose is medium in length and size, and his ears are of average size, aligned with his eyes. His jawline is rounded, and he appears to be a young adult.",
  "composition": "The image shows a close-up selfie taken from a low angle, with the user's face centered in the frame. The user is wearing over-ear headphones and a black hoodie. The user's expression is soft and approachable, with a slight smile."
}

Leading this LLM charge is yet again, OpenAI. The development of their AI technologies like GPT Vision has significantly altered the way we interact with our computers. This change is huge — it’s like stepping into the future. Now, instead of typing commands or clicking with a mouse, we can simply talk to our computers or show them a picture to get things done. This makes using computers much easier and more natural for everyone.

In the field of digital art, the progress is equally impressive. Tools like Pika and Runway are at the forefront of this change, enabling artists and even beginners to create stunning, realistic artwork with just a few clicks. Midjourney is yet another groundbreaking tool that’s making waves by producing images so lifelike, it’s hard to tell they’re created by a computer.

The Future with AI: More Connected and Intelligent Systems

As we look to the future, the exciting part is how different AI models are starting to connect and work together. Imagine having a team of smart tools, each good at different things, but when combined, they can solve bigger, more complex problems. This is not just a small step; it’s a giant leap in AI technology.

These connected systems can lead to smarter and more helpful AI. For example, in the future, you might have an AI assistant that not only understands your questions but can also suggest creative videos and images for your marketing campaign or even optimize your calendar.

This evolution in AI technology is more than just improving what computers can do; it’s about transforming them into partners that can help us in many aspects of our daily lives, from education to entertainment. As these AI systems become more integrated and advanced, we’re entering an era where the possibilities are endless, and the advancements we’re going to witness will be both remarkable and essential for our future.

AI in Existing Products

Generative AI in Mainstream Applications

Big companies are really embracing AI, making it a key part of the apps and software we use daily. Figma, a popular design tool, now includes an AI-powered design generator that can help create amazing graphics and layouts with less effort. Google and Microsoft are not far behind; they’ve introduced AI assistants, known as “copilots,” into their office software like Word, Excel, and PowerPoint. These AI copilots can help with writing, data analysis, and even making presentations. Apple, too, is enhancing its products with AI. A great example is the smarter autocorrect feature on iPhones and iPads, which makes typing messages faster and less prone to errors. All these examples show how AI is becoming a normal part of the software we use every day, making them smarter and easier to use.

The Emerging AI-First Approach in Product Design

We’re only just beginning to explore the full potential of AI in enhancing everyday products. More companies are starting to design their products with AI at the core, a trend known as the AI-first approach. This means that instead of adding AI as an afterthought, products are being built around AI capabilities from the start. This approach can make products much smarter, more intuitive, and more useful.

Here’s a nugget to chew on. Traditionally products had to be designed with a certain user flow in mind. Essentially you try to look at the happy path a user can take and build a code based interface around that. This is because once you write the code for a part of an app, it’s static, it can’t be changed until a whole new deployment is made. Now imagine a world where code can be written by an AI at runtime. If an intelligent AI was watching a user interact with a product and was dynamically updating the interface… you have a world where infinite user flows are possible. You no longer have to program for the mean case, you can program for goals and have the AI craft the experience to get to the goal in the most optimal flow for each individual user.

Development Tools

AI Changing How We Create Software

In the world of making software and apps, some really smart tools are showing how important AI is. We are immensely bullish on AI code creation tools. LLMs are naturally very good at programming. This is largely because they are trained on an immense amount of sample code, and if you think about it, C#, Python, HTML, are called programming languages for a reason.

Tools like Cursor, an AI-first Code Editor, Github Copilot, or Warp AI-terminal are dramatically reducing the amount of time and effort required to program. 41% of code being generated is created by AI!

How These Tools Are Changing Software Development

What does this all mean? Well, for one, the barrier to creating applications is dramatically going down. At the start of 2024, programmers are still necessary to actually build and deploy these applications. But the more these tools advance, it’s not unrealistic to imagine a reality where you can simply say something like “make me an app to track my calories, my daily weight updates, and my progress photos”, and within 5 minutes you have your own personal application, made just for you, and made exactly the way you prefer.

These tools aren’t just making small changes; they’re completely transforming how people make software and apps. They’re taking AI to the next level, making it more powerful and easier for everyone to use. This means that people who build apps and software can do things they couldn’t do before. They can solve harder problems, make their apps smarter, and even create new kinds of software that we haven’t seen yet. It’s a big leap forward, making the future of programming and app development really exciting

Hardware Innovations

AI-Enhanced Hardware Emerges

2023 marked the first year we started to see AI-first hardware products. Rewind and Humane dropped their first wearable AI pieces of technology. A close follow-up in early 2024 was Rabbit R1.

The key differences between these pieces of tech from all the other hardware advancements that came before is that they use AI natively. From their chips to their interaction design, they’re built from the ground-up with these superintelligent AI agents as a core focus.

The development of new gadgets with built-in AI, like Rewind’s AI pendant and Humane’s AI pin, is a big deal because it’s changing the way we use technology. These aren’t just regular gadgets; they’re like smart assistants you can wear. They can do things like help you remember important stuff or give you information right when you need it. This is important because it makes using AI super easy and part of our normal life.

Big companies are also creating special chips that make these gadgets even smarter. These chips are really important because they are the power behind the gadgets, making the AI work better and faster. This is a big step forward in technology because it means that AI can be included in all sorts of things we use every day, not just in computers or phones.

Why AI Gadgets Will Change Our Future

In the future, with gadgets like the Rabbit R1, we’re going to see a lot more devices that are made just for AI. This is really exciting because it means AI will be everywhere and help us in many ways. Imagine having a small gadget that can talk to you and help with your school projects, remind you of your friend’s birthday, or even give you directions when you’re lost. These gadgets could make our lives easier and more fun.

This shift towards AI-enhanced hardware is important because it will make advanced technology more available to everyone. It’s not just about having cool new gadgets; it’s about making our daily tasks easier and helping us to be more efficient. As these gadgets become more common, they will become a normal part of our lives, just like smartphones are today. This is why the development of AI-enhanced gadgets is such an important step towards a future where technology makes our lives better in many new and exciting ways.

Regulatory Landscape

Regulation in the AI Arena

One final point we wanted to touch upon is AI regulation. For most people in the AI world, regulation is a hotly discussed topic. Most people believe that AI regulation is necessary and some like Elon even believe we are behind the ball when it comes to regulating AI.

This camp of people have sound arguments: the rate of advancement is terrifying and the capabilities of these new models can be absolutely used for bad actors. Deep fakes, scams, and the like can have huge impacts on everyday people and without AI regulation, companies may unknowingly release tech that is too powerful and can be wielded with malice. In recent news, the US executive order limiting the publication of foundational model weights and the NYT vs. OpenAI case highlight the complexities in AI regulation. Additionally, with initiatives like the White House’s AI bill of rights, the future looks hopeful as we’re moving towards a more structured regulatory framework for AI.

Looking Ahead

And with that, we have arrived at the end of PressW’s 2023 AI recap.

From our perspective as AI enthusiasts, evangelists, and builders, 2023 was the biggest year of our lives. AI technology is growing almost exponentially and not a day goes by without a new tech, model, or announcement. I firmly believe that there is not a more exciting industry or time to be building. We’re very excited to continue pumping out LLM technology to our clients and helping redefine how businesses are built. We don’t anticipate any slowdowns in 2024 so stay tuned to get the latest updates into the world of AI. Thanks for reading everyone!

The State of AI 2023: How will LLMs redefine business? was originally published in PressW on Medium, where people are continuing the conversation by highlighting and responding to this story.

AI Readiness Check

So you want to start integrating AI and Machine Learning into your company’s products or workflow. Well you’ve come to the right place. In this article I’m going to break down some of the key questions to ask yourself to help you decide whether or not your company is even AI/ML ready in the first place. Hopefully this should save you not only time, but also money when you’re to start this journey, you won’t have pay someone to tell you the things I’ll be sharing in this article for free!

Define your problem and data

For any machine learning solution we have to start with the problem being solved and the data that is available. The problem and data are heavily inter-related. Data can often guide you to solutions you didn’t realize you could create and the problem you’re looking to solve will often dictate what kind of data is necessary. Picking one or the other should be your first priority.

I have a lot of data, what do I do with it?

If your company has a ton of data that has been collected for years, you’re probably wondering what is even possible with Machine Learning/AI, how can I effectively employ it?

If this is the case, you can start to ask what kind of problems does your data solve. Take for example you’re a house construction company. You’ve built 10,000 homes and on each you’ve collected house details like sq. footage, # of rooms, etc. as well as demographic data on where you’re building each home. Well with that kind of data you can very easily begin to imagine a product like a pricing sales engine that takes in all of that data and spits out what you should be pricing each home at to maximize revenue.

This is also an excellent time to bring on a ML/AI firm to help evaluate your data and work with you to figure out which custom solution best fits your business needs and objectives!

I have a problem I want to tackle, how do I set myself up?

In this scenario you have an exact product or solution in mind you’d like to build. Maybe that’s a database that supports natural english language search (semantic search), or maybe it’s that real estate pricing engine we were discussing previously.

Whatever the solution is, you must now walk backwards to figure out what data you need to solve that problem. A general good rule of thumb is, the more data the better.

If you want to build a recommendation engine for example, try to think about what data points would be most relevant to your outcome. Machine learning excels at finding hidden relationships between data points, but generally speaking, it’s a system that’s able to take what we intuitively know and generalize it across inputs. In plain english, the data is most obvious to us in terms of importance (number of rooms in a house dictating price), will most likely be highly important to the end solution.

Once you’ve figured out what data you need, the next step is just to start collecting! Collect as much data as humanely possible and take some time to setup your databases correctly to ensure it’s as easy to use for machine learning applications as possible. (Article coming soon on how to do this properly!)

Using a ML/AI Firm

Figuring out how to effectively deploy machine learning solutions can be tricky. Particularly when you factor in best practices around storing and collecting data, the custom nature of each solution, and the impact these solutions can have on your business.

I often recommend to business owners who are looking to break into this space, to actually just go ahead and hire an ML/AI consulting agency to do this initial setup and exploration. A lot of people believe that only after you properly setup your data, or have formulated what you want to build is the best time to bring on an agency like this, but I believe bringing in professionals as early as possible is the best way to set yourself up for long term success.

AI/ML agencies will generally do an initial exploration phase where they work with your business on figuring out exactly what kind of data is available and how to employ it most effectively. This usually comes with a data analysis component, which is a precursor the full ML/AI pipeline. These types of engagements generally cost around 25–50k to do and take place over a few months.

Conclusion + Shameless Plug

AI and Machine Learning are here and they are disrupting industries and businesses every day. It is in the best interest of every business owner to think about how to prepare themselves for this new wave of technology. This often starts with something as simple as ramping up the amount of data your business collects around its products/offerings. But I challenge you to really start to question where you can leverage this powerful technology within your business.

My agency, PressW, has been helping businesses answer these questions and begin to build their ML/AI futures from the day we opened our doors. We have helped businesses from healthcare, oil and gas, to construction, find unique ML solutions within their business that have been responsible for huge revenue and efficiency improvements within their business. We’re happy to sit down with you for a few hours (free of charge!) and discuss your business, your problems, and see if ML/AI is a right fit for you. If you or anyone you know would stand to benefit, please don’t hesitate to reach out to tarun@pressw.co!

Preparing your business for AI was originally published in PressW on Medium, where people are continuing the conversation by highlighting and responding to this story.

AI is molten hot, are you ready for it?

You’ve probably been living under some sort of naturally occurring aggregate of natural minerals (rock) if you haven’t seen the latest and greatest AI products shake up the Internet. GPT-3, ChatGPT, Stable Diffusion, Midjourney, and DALLE-2 emerged from the darkness and have dramatically shifted the way companies have begun to think about AI and employing it within their products and companies. If you’re someone looking to understand how to approach AI in your business/startup this is the article for you.

The Basics

One of the keys to understanding how AI and Machine Learning can be used within your company is to understand the fuel, the juice, the sauce that this technology runs on. One word, DATA.

Data is king in this space. It’s what fuels all these large machine learning models and allows them to do things like personalization, recommendations, answer questions, or tell you that that photo is indeed a dog.

One of the best first steps to assess your AI/ML readiness is to examine your business from a data lens. What kind of proprietary data is your business capturing? How is that data being stored? Is there additional data we can be collecting that could be useful later on. The general rule of thumb when it comes to that last point is that more data is definitely better than less. Storage is cheap nowadays, take advantage of that.

The Democratization of Data

In the past machine learning was only really capable in companies with vast amounts of data because they were the only ones with enough of it to train these models to become accurate. Proprietary data was king and if your company didn’t have its own data, you were often out of luck. You’d have to start by scraping the Internet for the data you were after, a time-intensive and costly operation. This is why for years companies like Google dominated the space, they had access to all the world’s data.

But while that still remains true, the best machine learning solutions are going to be found in companies with the most proprietary data, the releases of these LLMs (Large Language Models) are starting to paint a different picture. These models have already consumed vast amounts of data during their training, and since these models are now available on demand to the public, they have essentially leveled the playing field in terms of public data access. Now you can have something like ChatGPT understand the basics of cooking and put together recipes, without ever downloading a single recipe…truly spectacular.

But don’t be lulled here. It is still very important to be collecting propriety data in this age because that is the data that will fine-tune these large scale models to your particular use case. It still remains important to start collecting data as early as possible, storing it properly, and constantly be on the lookout for new data opportunities within your business.

The Importance of Starting Early

This will seem a bit, no duh, but I really want to emphasize this point. Starting early when it comes to AI and Machine Learning is very important. Not collecting and storing data efficiently, can have a huge cost on a business when the time comes to leverage one of these solutions.

Trust me, at our consulting agency PressW we have clients all the time who have not taken the proper steps to preparing their data. When they come to us they cannot wait to implement their new shiny sales engine or recommender system, but are often met with the disappointing reality that time and work has to be taken to get their systems in order well before we can even start doing any ML work.

Lastly here, to you as a business owner you should be considering your data as basically equivalent to money. Not collecting data properly, to me, is like letting free money walk out the door. You would never turn down free money, so don’t miss the opportunity to capture your proprietary data from as early on as possible.

In Conclusion

I hope from this article you’ve began to understand some key steps to take when considering AI and Machine Learning and their applications within your business. If there’s anything to take away I would say it is that data is your asset. Don’t let it go to waste and definitely spend some time setting up your systems properly to capture, retain, and manage as much of it as possible.

When you’re ready to supercharge your business with an AI solution or want to know if your business is AI/ML ready, contact me at tarun@pressw.co.

Guide to preparing your business for AI was originally published in PressW on Medium, where people are continuing the conversation by highlighting and responding to this story.

The Game is Evolving

What if I told you that there exists a technology that allows companies to build near native applications that can simultaneously deploy to not only iOS and Android, but also web, macOS, and Windows. You’d probably go “pshhh”, call me a liar, and click the little x in the top left and move on. That is, unless you were already familiar with Flutter, the powerful, open-source framework that is taking the development world by storm.

An ancient dilemma

For many years startup founders looking to build a mobile app faced an incredibly difficult choice when first starting out. Which platform were they going to build their product on? This seemingly simple question hides an incredibly deep trench of worrisome decisions.

If you wanted to build a mobile app you had basically two choices, prior to Flutter. Go fully native and face the burden of having to manage two separate code bases, deployment cycles, bugs. This essentially doubled the work, which usually resulted in a “Now available on iOS! Android coming soon” message on their website. Their only other solution was React Native, which had notoriously poor performance compared to native and was not efficient for complex interfaces and animations.

To add to all these issues, developing only a mobile app in this day and age made it difficult to reach users as a lot of users didn’t want to download yet another app, and would look for a web based solution instead. Meaning that founders would often have to develop for not one but three individual platforms to have full coverage!

Enter the Golden Child

In May 2017, Google released Flutter, a brand new framework that promised to whisk all these problems of the past away by allowing products to be deployed to iOS, Android, and web, all from one codebase. A “universal UI” so to speak.

Of course the world, myself included, were skeptical. Something that claimed to be that magical was bound to raise some red flags. But over time Flutter has proved to be a dominant force in the space of cross-platform frameworks. In 2022, it overtook React Native to become the most popular cross-platform framework in the world.

Flutter offers several technical advantages over its competitors. One of the main advantages is its fast development cycle, which allows for quicker iteration and prototyping. This is achieved through features such as “hot reload,” which allows developers to see code changes in real-time without having to manually rebuild the app. Additionally, Flutter’s built-in widgets and customizable design elements allow for a more consistent and expressive user interface. Flutter also uses the Dart programming language, which is easy to learn and has a simple, clean syntax. Furthermore, thanks to the way it compiles to machine code, it is able to achieve near native performance and buttery smooth animations.

Cost Savings and Reach

Using Flutter can also help startups and small businesses save money in the long run. Since Flutter is an open-source system, there are no licensing or subscription fees required. This means that businesses can create applications without having to pay for expensive software licenses. Additionally, the development time for Flutter applications is typically much faster than for other application frameworks, which can help businesses save on labor costs.

Additionally, thanks to Flutter’s cross-deployment capabilities all of a sudden startups are able to reach everyone all at once, leading to very powerful product experiences. No longer being limited in selection, with high quality performance across all three platforms, Flutter, in my humble opinion, has drastically shifted the landscape for new products.

My Experience

I felt that it wouldn’t be right to leave out the reasons why you should be listening to me. I’ve been using Flutter since 2019. When I founded my last company, Gathr (www.magical.app), we decided to take a gamble and pick Flutter over React Native, which was dominant at the time.

I’d like to happily report that the gamble paid off and I would say was one of the single best decisions we made in the product’s lifetime. We were able to use Flutter to simultaneously deploy that application to thousands of users, have an entire web experience, integrate complex machine learning recommender systems right alongside buttery smooth animations. It was truly glorious.

Since Gathr, I along with my team at PressW have been working startup founders and companies to produce Flutter applications for their businesses. We’ve received glowing feedback from the clients we work with because of how quickly their product was able to be developed, how high quality they look and feel, and how efficiently they were able to reach their audience. Many founders come to us asking for a web app because they believe it’s not possible to build a mobile app due to cost of development. Once we introduce them to Flutter, show them how we can get it not only on web, but also mobile, their entire company trajectory changes.

Conclusion

Flutter is quickly becoming the go-to solution for startups who want to build a product that reaches their target audience quickly and efficiently, while also saving on time and cost. In the past, there have been many limitations to developing a mobile app, but with Flutter, founders can now rest easy knowing that their product can be deployed to multiple platforms all from one codebase. It is no wonder why Flutter has become such a popular choice among startup founders, and why it will continue to dominate the development space in the years to come.

If you or anyone you know are interested in building a new product and are considering different tech stacks please feel free to reach out at tarun@pressw.co. Happy to give advice and see if Flutter is a good fit for your use case!

Flutter, your next mobile tech stack was originally published in PressW on Medium, where people are continuing the conversation by highlighting and responding to this story.

Introduction

With the recent advances in AI there has been an explosion of a new type of search engine, the semantic search engine. Unlike traditional search engines, semantic engines actually have an understanding of the words in the query and user intent (semantics) and generate search results using that understanding. Imagine you wanted to find the conversation you had with your friend about the best desserts you’ve had in your life. Today you might go into your phone and search “best dessert” and hope that one of you said those words. With semantic search, you can just type best desserts and the engine would be able to find your conversation you had with your friend about the amazing tiramisu you had in Italy because it understands that tiramisu is a dessert. Think about this technology applied to school notes, meeting transcripts, doctor notes, legal documents. The ability to just ask questions like we would to another person and immediately find the most relevant results is a massive step forward in this data age.

A brief history

Search engines have been around for decades and have enabled people to find the information they need with relative ease. They are arguably one of the most powerful and important pieces of technology in existence and allow us to sift through billions of webpages, as in the case of Google, or millions of products, such as on Amazon. Often people don’t consider how prevalent search engines truly are, but ask yourself the question: when was the last time you went to a webpage or used an app that didn’t have a search bar. A product today that served content would be nigh unusable with search.

But many of these search engines were built on keyword search, meaning when you type: “red shorts” into a site like Neiman Marcus for example, the search engine tries to match the words red and shorts to the product titles and descriptions in its database. It has absolutely no understanding of what the words “red” or “shorts” are or mean, to the search engine it was equivalent to asking for “slippery popsicle”. It’s just using the search words to directly match to its catalog.

Enter semantic search engines

In a simple example, our “red shorts” query to a traditional engine would probably not return us any “maroon” colored shorts, even though the colors are very similar because there is no similarity between the words red and maroon to a traditional search engine. However, our semantic engine understands that maroon and red are very “close” words because they stand for colors that are similar, and would thus return us some maroon colored shorts when we asked for red. To put it another way, semantic search engines not only consider the direct words themselves, they consider the intent behind the words, similar to how we as humans process words as we hear them.

Why is this a big deal?

I cannot stress to you enough about how important semantic search engines are. They are without a doubt the next evolution of search thanks to their ability to combine all the past factors like keywords with an understanding of intent. Semantic search engines understand what the user is getting at and can more accurately pull back relevant results. They unlock the ability to start searching vast amounts of data, but also unlock the ability to understand vast amounts of data. As we progress through this technological boom we start to see systems mimic human intelligence more closely.

While data has long since been an asset class in and of itself, the value of the asset is further increasing over time. Collecting data and converting it to text is no longer something that cutting edge companies are doing, it is becoming a standard, a necessity for companies to do, lest they fall behind their more data rich competitors.

What are the implications for the business world

There are already a number of industries that are and should be looking to employ semantic search engines in their operations. Any industry that has a ton of text based data to sift through are immediate first candidates, for example the healthcare, financial, and legal sectors. But today we are starting to see these new types of search engines enter into almost every industry from oil and gas, to education, and even e-commerce.

This really illustrates a larger shift in how data is being processed and understood. As we unlock new natural language processing capabilities the data that a business captures becomes more and more important because the tech can now extract more meaning from it than before. As a business owner it can revolutionize your internal operations by giving you the ability to search through every sales pitch your team does, search and understand every HR report, or search through all social media posts about your company far better than the built-in search engines those platforms support. Semantic search engines are the keystone to unlocking an entire new wave of business intelligence.

Conclusion

Semantic search engines are the search engines of the future. By understanding the context and meaning of search queries, they are able to return more accurate and relevant results. This enables users to find the information they need quickly and easily, resulting in improved user satisfaction.

If you are in an industry that deals with a lot of data and could benefit from better organization and searching capabilities of that data, I would highly recommend you start considering semantic search engines as a future part of your industry. They are an immediate and tangible boost to operations and can drastically improve the efficiency of many organizations.

My agency PressW has been building these semantic search engines for clients all across the US and in different industries. From our experience working in various fields like insurance, finance, healthcare, we’ve learned that each of these engines needs to be properly tuned and fitted to each client’s data. We know that he true value of the system shines when the data is properly prepared and integrated into the system. You have to take the time to really understand a client’s data, how it can be restructured to best serve these engines, and then the proper data cleaning and processing pipelines implemented to guarantee long term success.

If you or anyone you know are interested in building a powerful custom semantic search engine to power their business, don’t hesitate to reach out to me directly at tarun@pressw.co

Thanks for reading!

Note: Every image (minus the PressW logo) in this article were generated using Stable Diffusion on playground.ai

Why semantic search engines are the search engines of the future was originally published in PressW on Medium, where people are continuing the conversation by highlighting and responding to this story.

At Gathr, www.magical.app, we’ve been building our primary application in Flutter for over a year and a half now and we love it. Flutter allows us to quickly deploy our application to Android and iOS, and by utilizing Firebase as our BaaS (backend-as-a-service), we’ve been able to quickly scale up our product and feature set, thanks to Firebase + Flutter’s easy integrations.

Firebase is also exceptionally powerful for our specific use case as Gathr is a social-media-esq app, and we can utilize the Stream based Firestore queries to keep our app real-time for our events and chat features specifically.

The Issue

While incredibly powerful, performant, and scalable, Firebase’s Firestore comes with its own set of issues. Firstly came the problem of $$$. Firebase is a Pay-as-you-Go service, meaning that you are charged for your usage of their services past a certain free tier. Firestore in particular charges for every read and write that you perform. This means for a read-intensive feature, such as chat, the number of reads that you need to do, increases dramatically for every chat your user is apart of. Across scale, costs will run rampant as you now need to do document reads for each individual chat (group chats), as well as each chat message itself.

Without any caching, paging, or optimizations, let’s run through an example of the number of reads a user will trigger when loading their chat (Based on Gathr).

First some architecture details. Each individual chat in Gathr (group chats, DMs, etc) are stored as a document in a larger collection. Each one of these “chat” documents then has a sub-collection called “messages”, which stores any number of documents representing each individual message.

As you can see, for every chat that a user is apart of and opens, the number of reads goes up tremendously.

Let’s say a user is apart of 25 different chats, with each chat having on average 30 messages. Now because we have no local storage, just by opening the app the user triggers a read of at minimum 25 documents for each of their different chats, and then every chat they open costs an additional 30 reads. Just for opening one chat and reading the latest message, Gathr is billed for 55 reads! Now imagine that scaled across 100,000 users, with multiple chats receiving updates every day. The number of reads would be astronomical.

Obviously fetching every chat and every message in each chat is inefficient and can be pared down with pagination, we wanted to come up with a solution that allowed us to cut our reads down from total messages, to just new messages/updates.

We can do better.

Enter Drift

Drift is a SQLite DB written in Dart for Flutter applications, otherwise often termed as a form of “local storage”, an on-device database. Drift is going to help ease the burden of reads off Firestore by locally caching all of the chats and messages a user has.

Here lies an important question you may be wondering: if we’re storing everything locally, how will we know what messages and chats we need to get from Firestore and how will we update messages in Drift? Well okay maybe you weren’t wondering that, but you should be!

The answer lies in a DateTime timestamp. For every chat and chat message in our database we store a field called lastTouched. LastTouched is a timestamp that represents the last time any edits were made to that document. This is going to be incredibly helpful as we merge Firestore and Drift together.

We are going to setup our Drift database to mimic our Firestore database. We will create a table that stores a row for every chat that a user is apart of, and then another table that has a row for every chat message, where every message has a reference to the id of the chat it is apart of, so we can query later.

For a more in-depth tutorial on how to use Drift, refer to their docs here!

The Dawn of Lower Read Costs

So now that we have our Drift database setup to mimic Firestore, let’s go over how to save that $$$.

Our new chat system will work through Streams from our Drift DB. First we must establish a “Set Point”, which is the lastTouched timestamp of the last message in the chat (ordered by lastTouched). Then we can begin to setup our streams.

The first one will listen to all “older” messages, that have a lastTouched value that is less than (before) the Set Point. The second will listen to all messages in the database that have a lastTouched value greater than our Set Point. A separate Firestore stream can then be setup to fetch all new chat messages/updates to older messages (such as liking a message, editing the text, etc).

This may be a bit confusing to process so I’ll try to break it down a bit simpler. I’ll walk through a user experience on the app and cross reference what’s going on behind the scene.

A user boots their app and goes to a specific chat. First to note, all the chats are now locally stored so we have eliminated the 25 initial reads we were doing before to get all the chats they are apart of, woot! But we’re only halfway there.

Now when they load into a chat, they should see all their previous 30 messages (reading for free since it’s from Drift!). This is because as soon as they load into a chat, we fetched the latest message in that chat’s lastTouched field, and used that as our set point. We then opened two streams from our Drift DB. One that fetches all messages that have a lastTouched field before our set point (this contains the 30 messages they had prior) and one that fetches all after. One additional stream is setup from Firestore, which listens to the messages sub-collection and reports back any message documents that have a lastTouched field after the set point.

Now a new user sends a message, our Firestore stream dumps the new message doc into our database, and our second database streams reports it back out, showing the new message to our user!

AND BOOM!

Just like that,

we’ve cut the number of reads from 55 for a single chat, to just 1

Thanks!

And that concludes my crash-course on how to integrate Cloud Firestore with Drift DB to dramatically reduce the number of reads in an application. In our case we were looking at a chat application example, but this concept of using a Set Point + stream based setup should be able to work for any local storage system working with Firestore.

I hope you all got something out of this, please feel free to drop a question below, or reach out to me personally at tarun@gathr.team, and I’ll be happy to give you more insight or answer any burning questions.

Also check out www.magical.app to completely change the way you organize your social life :)

Flutter: How to dramatically decrease reads from Firestore using Local Storage was originally published in PressW on Medium, where people are continuing the conversation by highlighting and responding to this story.