Picture This: The benefits of rapid prototyping, and integrating AI in Miro

Generative AI is evolving rapidly. This makes it challenging for businesses to not only work out where to apply new generative AI technologies in the short-term, but also for their developers, who need to continuously upskill as new iterations of generative AI models, frameworks and products are released.

One of the ways in which companies can quickly understand new waves of generative AI technologies is through rapid prototyping. This involves quickly building working versions of products or tools that incorporate generative AI in some way, and serves several goals:

1. Learn about the technology at a deeper, more nuanced level, including how it works, what it takes to deploy the technology into a wider product, what it’s good at, what it’s bad at, what are its limits, etc...
2. Learn about what it takes to enable users to successfully interact with the technology, including what they use it for, how they interact with it, and what is needed for the technology to become a useful part of their daily workflow.
3. Potentially build an initial version of a product that could be scaled internally within the firm and/or for clients, although this is not necessarily the initial goal with brand new technologies.

This can be seen as the very early stages of agile product development, or even before product development: this is primarily about rapid learning and reducing technological and product uncertainty, which in turn will enable quicker, more focused downstream product development.

How we approach rapid prototyping

Here in Deloitte, we have several teams that focus on rapid prototyping, both to ensure Deloitte remains at the forefront of new technologies, and to help our clients do the same.

One of these teams is X Lab, which is part of Deloitte Digital.

In X Lab, we’ve been experimenting with AI for over a decade, including more recently with generative AI. We’ve been rapidly building generative AI prototypes both internally within Deloitte and externally for our clients, and in doing so, developed a much deeper and more nuanced understanding of the underlying technologies, and how they can deliver value within complex commercial organisations.

Our image generation Miro plug-in

One example of an early internal prototype that we developed was an image generation plug-in within Miro.

For those unfamiliar with Miro, broadly speaking it’s an online whiteboard tool, and one that’s used extensively in Deloitte for collaborative brainstorming and idea generation.

Our idea in a nutshell: could we integrate cutting-edge generative AI image generation into Miro so that our Deloitte colleagues could access this technology within their existing workflows? This would allow users to generate images within a single application, rather than having to combine Miro with other browser-based tools. This is would potentially only provide small uplift in productivity, but represented a quick and easy initial starting point to experiment with new image generation models.

The development process

We started with two workstreams in parallel:

1. User research to make sure we understood how best this technology could be leveraged to help our users. This information would be crucial in designing the app.
2. A scan of text-to-image AI models that would power this app.

We quickly found out that there were multiple AI models that could be used, including both proprietary and open-source models. We settled on an open-source model, as it provided sufficiently high-performance, and has a permissive licence for commercial applications.

We then designed and built a simple user interface that we connected to a Python backend that hosted the AI model. A key part of the development process was also building an architecture to host and serve the app securely and reliably, especially to handle any errors that arise, and to protect credentials like API keys.

Technical details

The architecture process flow was relatively simple:

1. We developed a front-end to serve the user interface on Miro.
2. This interface takes a text-based prompt from the user, and communicates it to a back-end that contains the image generation model.
3. The model takes this prompt and creates a corresponding image.
4. During this process, the front-end regularly verifies the status of the image generation process, similar to asking “Are you done yet?”.
5. As soon as the creation is complete, this image is retrieved and shown in Miro, and the user can admire their masterpiece.

A user flow of the app can be seen in Figure 1 below.

Figure 1. User flow of the AI Image Generator on Miro

Impact

After successful initial development and testing, the app was released to 32 separate Deloitte teams.

In the weeks following its release, we also followed up with teams to learn more about how they were using the app. Some were simply playing around with the tool as shown in figure 2. Some used it the tool to generate original images for their work as shown in figure 3. And some even discovered a new artistic passion, producing the result shown in figure 4.

Figure 2. Prompt: Magical Penguin

Figure 3. Prompt: Logo for a moodmeter app

Figure 4. Prompt: masterpiece, best quality, composition of human skulls, animals skulls, bones, rib-cage, jellyfish orchids and betta fish, (orange and blue bioluminiscent), intricate artwork by Tooth Wu and wlop and beeple. octane render, trending on artstation, greg rutkowski very coherent symmetrical artwork, cinematic, hyper realism, high detail, octane render, symmetrical, fan art, dramatic lighting, volumetric lighting, highly detailed, wind:1.2, 4k, 8k

Conclusion

This kind of app was almost unimaginable a few years ago, but now it is relatively quick and easy to pull together with the rapid advancement of image generation tools, including their availability as open-source models and their ease of integration into products. This will open up a huge amount of value in society by making image generation available to almost anyone.

This experiment also demonstrated that the value of generative AI is not only in the models themselves, but also in the products that are built on top of these models to allow these models to integrated generative AI more seamlessly into existing workflows. These products are essential in “unlocking” the value of generative AI for everyday users, and effective product development will be an essential part of realising the potential value of on-going advances in Generative AI.

For our team specifically, this experiment has provided an opportunity to see the different Generative AI tools out there, and has shed lights on the unknown unknowns in the field. Knowing these previously unknown variables enables us to estimate the amount of time and effort needed to incorporate Visual Generative AI into products. We, as Deloitte, are now in a much better position to advise on Generative AI strategies, having worked hands-on with the thing in question. As a matter of fact, we are already leveraging our learning to start work on an AI-powered Coach. That’s for an upcoming article though, so stay tuned and see you soon!