AI
“ When building models, customers don’ t want to deal with the platform, sourcing GPUs or building distributed systems,” Dwarak continues.“ That requires a lot of effort from the research team to go and optimise.”
This efficiency drive is closely tied to open-source foundation work. Rather than building proprietary silos from scratch, modern enterprise AI relies heavily on community-backed deep learning frameworks.
“ Many of my research team folks came from DeepSpeed, which is an open-source library that we contribute to,” Dwarak says.“ I myself have a background working on PyTorch... Many of the things we do, we give back to the community as well.”
This open-source foundation directly impacts how businesses handle their proprietary data. While many organisations start their AI journey by prompting off-theshelf closed models, enterprise workflows eventually demand domain specialisation.
“ A lot of our customers start with using frontier models,” Dwarak expands.“ It works fine for generic use cases, but then there are these business-specific user workflows where the data is very specific. Our customers are using open source to train their own models.”
A prime example of specialised deep learning over general-purpose language modelling is Text-to-SQL, which involves translating natural human inquiries into structured database queries. General foundation models often struggle with complex enterprise schemas because they lack access to real-time database context.
“Vanilla LLMs don’ t know how your data is organised. But we can look at the access logs and understand the context”
Dwarak Rajagopal Former Lead of AI Engineering and Research Snowflake
“ Text-to-SQL is a hard problem,” adds Dwarak.“ Vanilla LLMs don’ t know how your data is organised. But we can look at the access logs and understand the context. So the research team builds text-to-SQL models specifically post-trained on open-source models with reinforcement learning.”
What a successful agentic enterprise looks like As deep learning models become more specialised, the enterprise landscape is undergoing its next major evolution: moving from passive text generation to autonomous agentic architectures.
156 September 2026