mkdev's top 10 GenAI gifts of 2024

As we ramp up 2025, it’s important to reflect on the year that’s passed. 2024 was clearly the year of GenAI breakthroughs, topped with plenty of hype.

For many of us, GenAI has been a store of Red Rider sleds, enabling us to zip nimbly though terrain that before we could only trudge through. For some, however, GenAI may have seemed more like a piece of coal in the stocking, e.g. for companies who have paid big bucks for GenAI PoCs that never saw the light of production, or tech experts that were let go because their managers (naively) believed they could be replaced with AI.

Here’s the mkdev list of our personal 10 favorite uses of GenAI from 2024:

1. Fixing a customer’s GPU woes for their custom Large Language Model architecture

One of the highlights of 2024 was diving into the depths of a customer's custom LLM architecture -- a design so novel it made standard architectures look like they'd missed the innovation memo. But their models refused to cooperate with high-end GPUs, throwing bugs that could've been hieroglyphs for all the sense they made. We relished the opportunity to tackle the challenges posed by tensor mishaps and GPU quirks, and we dove headfirst into this exciting project. This project stood out because it wasn’t your run-of-the-mill "let’s call OpenAI’s API and call it a day" type of GenAI work. It was about getting down to the hardware and architecture level, where the real magic happens.

2. Automatically generating business documentation for a customer’s data warehouse

Writing documentation is like the broccoli of software development—necessary, but not everyone's favorite. That's why this project is a standout: we built a retrieval-augmented generation (RAG) tool to automate the drudgery. Instead of spending months poring over old code and missing crucial details, our tool magically retrieves all the relevant context for each component of a customer's data warehouse and feeds it into an LLM to produce top-notch documentation. The result was an incredible 30-fold speed increase, with our team only stepping in for a quick quality check. This wasn't just automation; it was automation done right. The custom retrieval system is like a magician, knowing exactly what to pull and what to leave behind, making it a total game-changer for efficiency and accuracy.

3. Generating a reference insurance pricing engine based on technical insurance documents for a customer’s IT migration project

Insurance IT migrations are the stuff of nightmares: pre-historic systems, regulator-approved documentation full of "tariff logic," and actuaries meticulously verifying calculations by hand. For one customer, the challenge was even greater - their ancient system couldn't produce the test cases needed to validate the new one. Enter GenAI.

We built a cutting-edge reference pricing engine by converting technical insurance specifications directly into Python code using an LLM. But we didn't stop there. We turbocharged quality assurance by extracting test cases from the same documents and even developing a "delta-explainer," which pinpointed discrepancies between the reference engine and the new system, tying each issue back to specific parts of the documentation. This project was a massive time-saver, and it showcased the real power of GenAI when paired with tailored engineering. Off-the-shelf tools just couldn't keep up with the complexity, and even prompt engineering could only take us so far. The real magic came from crafting custom solutions that bridged the gap between regulatory precision and cutting-edge automation.

4. Automatically refactoring code with automated quality assurance for a customer’s data warehouse

Refactoring code is like cleaning out a storage closet—necessary, but nobody's rushing to volunteer. For one customer's data warehouse, we tackled this chore head-on using GenAI. A lot has been written about using GenAI for code generation, with results largely, though not universally, positive. We decided to put it to the test, feeding an LLM both the code to be refactored and examples of code written in the target design.

While off-the-shelf GenAI didn't immediately excel, we rolled up our sleeves and embraced the challenge. We crafted a custom workflow that automated retries when the refactored code didn't pass tests, sealing another triumph for automated testing. We also implemented a semi-automated process to handle functional differences in the outputs, as random processes in the original code made a perfect match unnecessary. This project stands out because it transformed one of the least-loved developer tasks into a seamless process, showcasing the incredible power of combining GenAI with thoughtful engineering.

5. Autogenerating image tags for our website

Tagging images on a website is one of those small, unglamorous tasks that quickly snowballs into a monumental time sink when you're dealing with hundreds (or thousands) of files. For our website, we decided to let GenAI take the reins, and boy, did it deliver! Using GPT-4 Vision, we automated the generation of descriptive alt text and tags for over 1,000 images. The setup was a breeze, and the results were nothing short of amazing. GPT-4 Vision analyzed each image, generating detailed descriptions and then translating them into concise, accurate tags. To ensure relevance and alignment with our SEO goals and accessibility standards, we fine-tuned the output. What could have taken weeks was done in hours, and the results spoke for themselves: a fully tagged image library that boosted both usability and search rankings. It's truly remarkable how we transformed a once tedious task into a seamless, GenAI-driven process.

6. Making cute Pomeranian movies

So far, all the cases in our top have been serious, but we always make time to have some fun. When we got our hands on Sora, the first video we generated had to be of Pomeranians. It would’ve been ridiculous to start with anything else (especially since our co-founder is a proud Pomeranian owner).

You can definitely tell the video was generated, not only because it’s a well-known fact that Pomeranians can only drink from cups with a single handle, but also because, with food just 20 centimeters away from the Pomeranian, it’s extremely unnatural that it hasn't been eaten yet.

7. Demystifying vector embeddings and databases

Picking the right database for AI is super important for projects, and it's a decision that can make or break things. For us, it was all about diving into vector databases and figuring out how they could improve AI workflows, especially when dealing with unstructured data like text and images. Using GenAI, we were able to streamline the whole process by automating the decision-making process about which database is best for each AI task. It's a perfect example of how GenAI can make decision-making easier, cut out the guesswork, and make AI infrastructure smarter — and without all the usual hassle.

8. Generating customized lead emails

As we've shown in previous rankings, GenAI often serves as a helpful intermediary rather than the end product itself. So, we decided to experiment a bit and put it to work in our very own sales and marketing outreach communications. With the power of code and webhooks we built a custom tool that pulls customer data from open sources, then uses language models to structure it just the way we need it. From there, GenAI generates a lead email, following a set of rules we created. These rules take into account things like company activities, founding dates, team size, technologies used, and even our own services, competitive advantages, and ideal customer profiles. The process was almost 100% automated, all you had to do was feed the lead's contacts and wait for a response.

It's an incredibly handy automation tool, but let's be clear -- it doesn’t replace the personal touch that’s vital for building relationships in our consulting business. Still, as we had fun building and testing it, we couldn't help but realize that the future of GenAI in marketing and sales is looking pretty bright!

9. Contributing to a national EU AI Act training curriculum

The EU's AI Act, which was passed in 2024 and is set to enforce compliance requirements in 2025, brought a new wave of regulatory challenges for AI businesses. One country's innovation office reached out to us to help shape a training curriculum for AI Act compliance, and we were so excited to jump in!

The fast growth of GenAI was a big part of the changes made to the AI Act between its 2021 draft and the final approved version, so understanding these changes was really important. Paul, our Head of Data + AI, who was in charge of data science governance at Allianz, used his experience with AI Act compliance, as well as his background in quantitative risk management. His past work included running workshops on risk and AI and engaging directly with regulators on both quantitative and qualitative risk topics, making his insights vital for this effort.

10. Indexing our personal knowledge stores to semantic search (almost) everything

In today's world, where there's always more and more data, it's more important than ever to manage and access information efficiently, right? We decided to take on the challenge of indexing our personal knowledge stores and integrating them with semantic search capabilities. By using vector databases (and a little help from GenAI), we made it possible to search through a wide variety of resources and retrieve exactly what we need, almost like magic. Whether you're looking for documentation, research, or past projects, this system helps you find the information you need based on meaning, not just keywords. It's a great example of how AI can make searching and organizing data more efficient and smarter. We're still making some tweaks, but we can already see how this setup will completely change the way we interact with our internal knowledge in the very near future. With GenAI by our side, we're building a new kind of searchable intelligence just for ourselves! And guess what? We already have more ideas for 2025…

As we wrap up this loosely structured ranking of our favorite GenAI experiments of 2024 (it can be read in any order, it's not necessarily sorted from most favorite to least favorite), we see that the potential of AI is enormous. We experienced it firsthand with our clients. From improving internal workflows to enhancing marketing strategies, GenAI has proven itself to be a powerful tool that speeds up processes and opens up new ways of doing things. But of course, there are some things that AI simply can't replace.

What will never replace GenAI? Rina's amazing graphics. Rina is an incredibly talented artist who creates all of our illustrations, and while we tried to get GenAI to emulate her style, let's just say.... it failed miserably. The result was, well, a bit disastrous. Rina's unique artistic writing is something no AI can replicate.

If you'd like to see Rina's stunning artwork for yourself, why not check out our merchandise store? You can purchase some of her incredible designs, from t-shirts to mugs. Trust us, it's well worth it!

And speaking of things that can’t be replaced – if your business needs custom solutions, cloud optimization, or just need to understand how GenAI can work for you, we’re here to help. Let’s make 2025 even more exciting together!