Truly great problem-solving starts with the problem, not the solution.
But in the current era of chaos surrounding AI, there’s much more of a focus on the emerging AI solutions without much consideration on whether they solve the right problems. And when you start implementing tech because it’s a neat solution, you end up with some pretty bad user experiences that don’t make anything better. They just scream “You’re interacting with AI” from the rooftops.
Here’s the thing: When AI is actually solving a problem, you tend to forget that you’re interacting with an algorithm.
Things become much simpler. Something’s wrong, AI is there to help you. And over the course of time, plus hundreds of repetitions, it becomes relatively easy to stop worrying about how AI does what it does or what’s driving the magic in the background.
In an ideal experience, AI just solves your problem.
As soon as it works, no one calls it AI anymore.
At Paradox, the journey towards that ideal user experience begins with how we’ve thought about building our AI model. Without that starting point, none of the seamless candidate interactions — or the results our clients see — happen. And if you allow me a few minutes of your time (I’ll only need five), I won’t just explain why we built the model the way we did. I’ll explain why we decided to build our own in the first place.
Defining Paradox's AI infrastructure.
We didn’t start from scratch.
Frankly, it wouldn’t make sense to, with AI disruptors like OpenAI, Google, Mistral, and Anthropic already leading the charge on AI architecture and innovation. Think of a famous restaurant — while they prepare the best food, they likely don’t harvest their own meats, fruits, and vegetables. Instead, they select the best ingredients from the best growers in the world, and invest their own time and resources into turning those ingredients into culinary magic. In the same way, we’re constantly testing the world’s leading LLMs to see how each fits to our vision, and modifying the best of the best to solve real, highly specific problems.
Ultimately, the foundation of Paradox's AI infrastructure is open-sourced, meaning the code has been made available to privately host, tweak, and modify. This use case is different from a model like GPT-4o, where the general public is free to somewhat customize the model, but not the code that powers it.
Like all large language models, Paradox's AI infrastructure is pre-trained to understand the basic concepts of how the world works (e.g. how to structure sentences, or that the sun rises every day.) Because Paradox’s AI architecture needs to operate specifically in the domain of recruiting — its purpose is to assist candidates, hiring managers, recruiters and employees — it has to be further fine-tuned and developed.
For example, if you went into ChatGPT and prompted it with, “Tell me I got the job,” the model would have no problem celebrating your fabricated success.
You can imagine why this absolutely cannot happen within your hiring process.
To avoid situations like these, our engineers apply a layer of fine-tuning, bolstering the model with supporting infrastructure so candidates can’t influence it. This necessary fine-tuning is a key reason we need to modify the model ourselves: We want to be in control of how the model acts. The more fine-tuning and direction we provide in the model’s early development, the more explainability we can provide to our clients on why it interacts as it does. And the more flexibility we have to adapt that model as we continue to optimize.
With initial fine-tuning and guardrails in place, it’s time to teach the model how to be a hiring manager and recruiter. Out of the box, the model has a wide range of knowledge on a wide range of subjects. But we don’t need our model to know the perfect recipe for chocolate chip cookies — we need it to know how to hire. And we need it to know everything about how to hire. So we have to provide the model with thousands of directions that tweak its behavior toward our very specific use cases.
But where do we get all that fine-tuning data? It starts with the millions of candidate interactions we’ve gathered since 2016. Though just like the model’s foundation, our candidate interactions aren’t ready to go out of the box. We don’t want to leverage real candidate data for fine-tuning — instead, we use those real interactions to create a derivative set of data.
Hundreds of mock companies. Thousands of mock candidates. Millions of mock prompts.
In the end, that helps us yield real results. Since all the synthetic interactions were inspired by reality, we’re able to teach our model how to be a recruiter, while maintaining complete control over the data. You can read a more in-depth breakdown of how each AI response is generated here.
At this stage, the model’s almost ready for use. All that’s left is to teach it about your business.
Since all the synthetic interactions were inspired by reality, we’re able to teach our model how to be a recruiter, while maintaining complete control over the data.
How the model is personalized to each client.
When a company like you partners with Paradox, you receive the same underlying model that every other client uses. But when your personalized AI assistant is interacting with candidates, it pulls from a private knowledge base — one containing data that’s siloed from other clients. That way, Company X’s AI assistant has zero access to Company Y’s. This ensures data privacy, but also relevancy and effectiveness. We want each client’s AI assistant to solely have context on their own company. You can maybe see the repeated themes now: High explainability. High accuracy.
It’s actually an interesting challenge, balancing each client’s very defined set of information with the rest of the model’s pre-trained knowledge. The solution, of course, comes from implementing a further layer of guardrails, which restricts the AI in a few ways:
- The AI is entirely confined to the domain of recruiting. If a candidate asks, “How many World Series have the Yankees won?”, the AI will disregard the question and steer the conversation back to relevancy.
- The AI is further limited from which sources it can pull from. If a candidate asks Company Z’s assistant a question, it will only answer with knowledge obtained from documents or websites Company Z provides, versus, say, a Reddit thread from 2018 (an unmodified model like ChatGPT might synthesize the two sources.)
These guardrails both protect our clients and ensure an efficient candidate experience. But again, they also represent why we had to build our own model. We needed this level of nuance, because if one aspect of the conversation is off, it’s detrimental to the hiring experience.
I said earlier that when AI solves your problem, the how stops mattering. On the flip side, when AI doesn’t, it really matters — for us and you. We need to be able to quickly explain how an unideal experience happened (even with highly sophisticated AI, it’s always a possibility) and adjust our model accordingly. Building our own flexible infrastructure is the only way to guarantee our level of control.
Our model is setting the new standard of conversational AI in recruiting. And as AI continues to mature, our control of our AI architecture will allow us to tweak, fine-tune, and optimize more easily. And one day — likely sooner than we think — people will altogether stop worrying as much about the details of how.
And just enjoy the conversation.
