It has always been my goal and dream to be a researcher. It’s what I think I’m best at and what I enjoy doing the most. When I came across Cohere’s Scholars Program, I knew it would be something I would love doing, so I applied.
Unfortunately, I did not make it past the initial round. And while that was definitely disappointing, I still believe I would have had a lot of offer to Cohere, and since I put a great amount of effort into my application, I wanted to share it here.
The application consists of two things:
- A 5 minute video discussing possible extensions to one of Cohere’s recent research papers, which you can find here.
- A personal statement in under 500 words, which you can read below.
Personal Statement
29 August 2025
I like to think that I have always been a researcher at heart, but just haven’t found the right role for it yet. This is why I am so excited to be applying to the Cohere Labs Scholars Program. As an example, I have always had a strong intuition for what is (or will become) relevant. At Microsoft, I was successfully applying retrieval augmented few-shot learning back in time when OpenAI’s “text-davinci” model was still the best in class. My passion for research and teaching also extends to our internal reading groups, where I found I could not only hold my own with senior engineers and scientists, but often teach them something new. Presenting the LoRA paper just days after its publication is just one of many such examples. However, my research ambitions are much bigger than that. I want to create new research.
In fact, thinking about open problems in ML has led me to pursue my own, independent research project on “CCLIP: Conditional Contrastive Language-Image Pre-Training,” which rests on the idea of shared image concepts as an additional gradient force. By using masked captions, hard negatives, and a SigLIP-style loss, I believe this approach could mean a quantum leap in training efficiency, in line with one of Cohere’s research priorities. Even more excitingly, a successful CCLIP model itself could serve as a backbone for future iterations of Command or Aya Vision. However, while the project allows me to build upon ML research code, it is also a challenge to do so all by myself. To transition from idea to a scalable model, I would love to have the research environment and mentorship that the Scholars Program provides.
As a data scientist, I thrive on turning vague ideas into useful products and have done so numerous times at Microsoft – from RAG pipelines that surpassed a strong human benchmark to widely used AI automation tools. At Skyscanner, I have already transformed how we think about search and ranking by driving the use of cross-encoders. As a testament to my written and verbal communication skills, I would also like to highlight my experience teaching ML while working part-time at LSE, and my blog posts, which have been read by hundreds of thousands. At the same time, I have also immensely advanced my technical skills. Today, I am writing thoughtful, well-designed code other developers love using. However, I want to go much further and learn from Cohere’s experts to solve the modern engineering challenges of large-scale, distributed model training runs.
I cannot think of anything more exciting than collaborating with some of the brightest minds on novel ML research ideas and publications, and my goal is to continue this work after the program as a research engineer. I fully believe in AI’s vast positive potential, but also in our responsibility to distribute its benefits fairly and widely across the globe. This is why open research and open source play such a critical role, and Cohere’s leadership in this space is incredibly inspiring.
Andreas Stöffelbauer