1
The rich text element allows you to create and format headings, paragraphs, blockquotes, images, and video all in one place instead of having to add and format them individually. Just double-click and easily create content.
A rich text element can be used with static or dynamic content. For static content, just drop it into any page and begin editing. For dynamic content, add a rich text field to any collection and then connect a rich text element to that field in the settings panel. Voila!
Headings, paragraphs, blockquotes, figures, images, and figure captions can all be styled after a class is added to the rich text element using the "When inside of" nested selector system.
It’s difficult to find a modern organization that isn’t taking a hard look at artificial intelligence. Not only is it all over the news, but new advancements are unlocking capabilities that weren’t possible even five years ago. It was unthinkable to outsource customer support to AI in 2020 - now, it’s becoming the default.
Going from idea to production isn’t nearly as straightforward as, say, a software engineering project. This is uncharted territory for most companies, and few have the expertise to handle it themselves. AI solutions sit at the crossroads of engineering and data science - bringing with them, the problems of each.
There’s a lot to consider - building out a mature in-house data science practice is not easy.
Instead of spending years figuring all that out, a vetted team of experts can get you to a solution quicker. AI agencies (like ours, NineTwoThree) bring the experience, you bring the problem and the data.
It’s important to understand what you’re signing up for, though.
If you’re at the point where you’re considering an ML project for your company, then congratulations - you’re getting in at the right time. This technology is going to change the ways a lot of people interact with the world. But if there are two words to sum up the AI hype, it’s “unrealistic expectations.”
There are so many considerations at every stage of the process - data, model selection, training, fine-tuning, testing, bringing to production…each stage brings complications and delays. And if you’re partnering with an agency, you’re going to want to make sure things are on-time - otherwise, costs can add up, and projects can sit idle.
Let’s go over each stage of the process, and understand what it takes to deliver on-time.
It truly takes a village to make a world-class AI solution.
You’ll need expertise across data science, software engineering, DevOps, legal, and product…and that’s if the app DOESN’T face any customers.
A few common scenarios:
Understanding these gaps is key, and step one.
“If you’re building an AI solution for the restaurant industry for example, it’s important everyone has a base level of domain knowledge. What features are important for restaurants? What influences their revenue? Getting this information early can save time later.” says Vitalijus Cernej, a Machine Learning Engineer at NineTwoThree.
Trust us, the sentence “we’ll handle it ourselves” should only come up if you’re CERTAIN the expertise is there.
One is reminded of Hofstadter’s Law:
“It always takes longer than you expect, even when you take into account Hofstadter's Law.”
Any project is going to have unpredictability in the timeline, but that doesn’t mean there shouldn’t be expectations.
“Experience is the cornerstone of efficiency. The real question isn't if a problem will arise, but when.” says Vitalijus.
This is where leaning on the experts can help. They know what usually takes longer, what usually takes shorter, and where risk occurs. Scope creep, integration issues, resource allocation - all common problems to run into. They’ll also know what to do when a problem occurs, and how to address them quickly.
A great first step is to just come to an estimate, even if it includes variability. A well-defined statement of work can solve for this, and real-time visibility using project management tools like Monday helps keep expectations aligned.
The inevitable next question is…what happens if delays happen? What happens if BIG delays happen?
This is another issue you can agree to tackle when it comes up. Regular communication should be par for the course, so discuss these in your Slack channel or during your weekly call.
There are a few ways to concretely handle this, though, which we’ll talk about next.
If you’re reading this, you’re probably not Facebook or OpenAI, who devote billions of dollars to train their own AI models from scratch.
So, at some point, you’ll be working with someone else’s solution - be it cloud compute, AI models, or data lakes.
But there’s a layer right below that where you might run into issues trying to build your own solution. It could be at the start, realizing that adding functionality in a key component balloons cost or development time.
Or it could be during the project - you’ve been working on that component for a few weeks, and are realizing it will take more $$$ to get it working than you’d like to commit.
Either way, the discussion of outsourcing technical solutions inevitably comes up. There needs to be a conversation - discussing tradeoffs of continuing to develop your own solution vs outsourcing.
Answer questions around:
This can help you make decisions about getting unstuck later on.
Successful AI projects are narrow in scope. Instead of asking an AI solution to handle everything, define specifically what you want it to do. Going from broad panacea to “offload this step in the process to AI” is how the discussion should start.
“The expectation is that the LLM is just like a person, and can do everything super quickly. But expecting that much out of a system that gets expensive - if you focus on what’s necessary, then the system can still perform really well. The idea is to simplify.” says Jurgis Samaitis, an MLOps Engineer at NineTwoThree.
This is especially true if the idea started as a flashy hackathon project or demo. Those are great as a proof-of-concept, but a working demo != a working project, and setting expectations at the start helps everyone understand what success looks like.
Define what process or job the AI will perform. Especially if this solution is customer facing - if any AI-generated output will reach the customer, there need to be guardrails established to make sure it represents the brand well.
Answer questions like:
Better yet, write down the solution scope at the beginning. This initial step is crucial. In fact, knowing WHEN to use an LLM is half the battle.
Taking a smaller, modular approach - where the LLM is inserted at a specific point, for a specific purpose, with safety if something fails - is a great way to start.
It’s almost impossible to overstate how important data quality is to any successful AI project.
“You expect that the data is going to be clean, informational, and quality, but in nine out of ten cases, that’s not happening.” says Vitalijus.
The old saying goes “garbage in = garbage out”, and that’s no different here. The best-designed system, with incredible feedback loops and robust safeguards, will still perform terribly if the underlying data has issues.
There’s a few things to consider when it comes to quality data:
“You need proper data pipelines as well,” says Vitalijus, “and you need to figure out extract-transform-load (ETL) processes first, before you even start training.”
It’s important to know what a good data pipeline is. To define that, let’s go over what “bad” looks like:
If you spend most of your time on this step, that’s not necessarily a bad thing.
A word of warning: this can feel like a tarpit for any machine learning project. It can take longer than you’d expect to figure out your data problems (and that’s why billion-dollar companies exist to solve that specific problem.)
However, if you feel like you’re spending too much time on this step, you probably aren’t - that’s how critical it is.
Sort of. Most people think that which model you pick determines the outcome of the entire project. So they want whatever’s the best, and usually the most expensive. Think the latest Claude or GPT. This might be the best answer, but it comes at the cost of speed and price.
So sometimes, that’s helpful. But there are plenty of other models out there.
We’ve found that picking the perfect model for the first pass through fine-tuning isn’t necessary. Great, cheap models like GPT-3.5 can teach us a lot about the performance and feature selection, without breaking the bank.
“Data is king, and feature engineering is king. If you’ve prepared your data in the right way, you can get to 80-90% accuracy using a cheap base model. That’s more than enough for a proof-of-concept.” says Vitalijus.
Yes, down the road, it’s better to dial in a model. Maybe a model trained specifically on code, if you’re working on AI that’s generating code.
But for testing at the start, speed and quick iteration is the most important consideration.
This is usually where a lot of friction and delay can happen. Most people skip over this in the planning phase, and expectations are misaligned on what “correct” looks like.
Think about it - most software projects have predictable outputs and specific outcomes.
“One of the key points in terms of timely delivery is to remember that this is statistics. It’s predicting the next word, and it’s really good at that. But statistics is unpredictable in nature, and there are gray areas.” says Jurgis.
Statistics can be unpredictable - it’s a really good approximation, not a guarantee of an answer. So when you create a chatbot to answer questions, and the answer isn’t exactly the same each time, it can be surprising. It’s important to understand that fixing one response from an LLM might make ten other responses unreliable.
This is where it’s helpful to bring in the experts on your team. If you’re replacing a manual task, talk to the employees who won’t have to do the task anymore. They’re best-equipped to understand what “correct” and “adequate” look like in terms of performance
Agreeing on an evaluation suite can clear up confusion and communication cycles, too.
“When you have a lot of subsystems, and something hallucinates, you’re not going to be aware something went wrong until you have very bad feedback from the customer. So it’s important to have robust evaluations at the start, to compare the output of every iteration.” says Vitalijus.
If you didn’t do this earlier, do it as soon as possible now.
We speak from experience - there’s a point where adjustments hit diminishing returns on these projects.
R&D improvements in machine learning are not linear. Getting from 0% to 95% complete can often be straightforward, but it can take just as long to get from 95% to 98% (and forget about 100%.)
“Achieving 100% accuracy is impossible for most ML use cases. We live in a non-deterministic world, and while data reveals patterns, there are almost always outliers and unexpected events that impact predictions.” says Vitalijus.
Some PhD-level professionals spend years improving a single metric by a few percentage points - it’s that complex and difficult to do for non-deterministic ML products.
So, it’s important to define what “done” looks like, and revisit that definition often. If you’re building an internal tool to help your team, it’s likely already offloading a significant chunk of their work. That means it probably shouldn’t be perfect 100% of the time.
That’s a dramatic difference from a public-facing tool that customers will use - they’re expecting an extension of the company, and performance and accuracy should reflect that.
Work with your team and the agency to determine what’s realistic for a complete project. And we truly mean realistic - 100% accuracy is extremely difficult to achieve for AI projects.
That also means defining what ISN’T good enough. If the entire system hinges on doing one thing really well, then it should perform that one function extremely well. Standards should remain high, and that can reflect in the testing and integration phase.
It’d be irresponsible not to address this. There are two main parts to liability.
The first is ensuring your data is secure, with strict governance and access controls. This should be table stakes for any modern enterprise, but it’s especially crucial for sensitive data, like healthcare information or legal documents.
Second is guardrails for data being exposed to the outside world. This is something to agree on early in the project. Is it a liability? If so, how much should we invest in ensuring it doesn’t happen? That affects your testing and design phase considerably.
Spending more time, exploring solutions, upgrading models, testing…all adds considerable cost to any project. And if you have infinite money to throw at this, then this section isn’t for you.
But if you’re focused on building a sustainable business built on AI features, cost should be your number one concern. It’s easy to see costs balloon when chasing the last percentage of accuracy improvements, or upgrading models from GPT-4 to GPT-5.
This section could be its own article, but the main gist is simple: following the advice in the previous sections helps control cost. Increasing predictability, limiting scope, having upfront conversations to align everyone…that creates a nimble, streamlined project.
And if it does come time to adjust, make it a team decision.
But sometimes, it’s helpful to see a real-world example.
NineTwoThree was selected by the CR Innovation Lab to help build an experimental chatbot that combines the power of AI with CR's expertise to answer your questions and offer product recommendations.
There’s a lot to consider if you want an efficient project.
Partnering with experts is an excellent way to get a head start on your organization’s AI journey. Don’t let unrealistic expectations get in your way - communicate often, align on every step, and you’ll be well on your way.
In case it isn’t clear, we’ve got a lot of experience with this. If you’re interested in how AI can save your organization real money and time, let’s chat.
