Talk about your next project
Nebius case study

The lesson

When your buyer is an engineer, tell the customer story as a technical write-up, with each section explaining one technique the customer used and what it gained them.

The move

The title "YerevaNN: LLMs for molecular optimization" sits beside a lab photo, followed by three cards: a "Premise" summary of the research (three models pre-trained on 110M molecules, 40B tokens), a profile of YerevaNN as a non-profit research center founded in 2016, and a "Contents" list of six technical sections. Stat tiles and a summary paragraph follow, reporting an 8% improvement over previous methods and processing rates of 180,000 words per second, backed by a small "Words per Second" table by batch size. The body then works through "Flash Attention, compute capability and custom Torch extensions" and "Environment management", explaining each tool before stating YerevaNN's choice, such as Conda-Pack for 30% faster environment porting times on Nebius H100 systems.

Why it works

  • Our readThe Premise and profile cards tell an ML engineer who the customer is and what they trained before any infrastructure detail appears.
  • Our readThe Contents list names the techniques up front, so a reader can jump to the one they are fighting with, such as FSDP or DataLoaders.
  • Our readEach section explains a tool and its tradeoffs before naming the customer's choice and its gain, so the result reads as an engineering decision rather than a sales claim.

What not to copy

Don't copy YerevaNN's molecular drug design context, the Flash Attention and Conda-Pack specifics, or the research conference tiles. Keep the premise, customer profile and contents cards under the title, the stat tiles with a short results summary, and the body built as one technique per section ending in the customer's decision and outcome.

More from Nebius

More case studies

Steal it for your brand.

Tell us what you want to make. We’ll talk through how to put these ideas to work for your brand.