Ritwik Raj

Stop letting probabilistic LLMs directly dictate your transactional workflows. Over the last few months of building and scaling agentic systems, I've noticed a fundamental flaw in how many teams approach AI integration. They treat the LLM as the orchestrator. But large language models are inherently non-deterministic. If you are building enterprise software, your backend execution—especially when orchestrating complex sagas or updating document stores—demands 100% determinism. The solution is establishing a strict "Determinism Boundary." Here is how we need to architect these systems: The Probabilistic Zone: Use the LLM (like the Gemini API) strictly for intent classification, reasoning, and data extraction. Let it handle the messy, unstructured human inputs. The Boundary (Constrained Decoding): Never let the AI output raw text to a service. Force the output into a strictly typed state object (like a JSON schema or a strongly-typed Java record) using constrained decoding and tool calling. The Deterministic Zone: Once that structured payload crosses the boundary, the AI's job is done. Your traditional deterministic architecture—whether that is a Spring State Machine, a distributed saga pattern, or a robust API gateway—takes over the actual execution and database mutations. The AI decides what needs to happen. The engineered system dictates how it happens safely. If you try to make the LLM do both, you are just engineering future outages. How is your team handling state management and execution in AI-driven applications? Are you using specific frameworks to enforce this boundary? Let’s discuss in the comments. 👇

20

 New Blog Alert: “AI Duct Tape Isn’t a Strategy” I finally wrote the piece I’ve been asked about a lot recently (and occasionally guilt-tripped into writing): the difference between AI-Enhanced products and AI-Native ones. (Please stop taping AI onto things like it’s arts and crafts night.) If you’ve ever sat in a meeting where someone suggested “Let’s just add AI here” with the same enthusiasm as putting a sticker on a laptop, then you’ll enjoy this one. I cover the architectural, mindset, and organizational shifts that are required, and why slapping AI onto a legacy workflow is the tech version of taping a smartwatch to a sundial. 👉 Read the blog: "AI Duct Tape" Isn’t an AI Strategy https://lnkd.in/e7aSjyaX If it makes you laugh, share it. If it makes you nervous, you should definitely read it. If it makes you re-architect your roadmap… you’re welcome.

8

2 Comments

This is one of the most insightful posts to come out of GTC for enterprise AI. Connect these ‘data’ dots to what economic diffusion means to the speed of “business value” from Data+AI. In real life the CuDF and CuVS layers accelerate the lowest layers of data organization, speeding up data lookups. The upper layers of data are based on semantic modeling, knowledge extraction, context graphs and agent memory. Tokens will help organize this better, speed it up; but there’s a lot more to be done to map a disparate data-driven application to become agentic. Agent factories with rich agent runtimes will plug this gap - tokens (from AI factories) and data in, business outcomes out - turning trusted data into (autonomous) trusted action. Data gravity leads to Agentic gravity. Allen Hosler Kumar G. Varun Sandesh Rao Arjun Upadhyay Ron C. Sungpack Hong Rhicheek Patra, Ph.D. Ajit M.

17

Excited that Cortex Code CLI is now generally available. Cortex Code is the Snowflake-native AI coding agent that turns complex data workflows into simple interactions—all in natural language. Whether you’re in Snowsight or working via the CLI, Cortex Code is deeply aware of your Snowflake environment, meaning Snowflake expertise becomes an always-on capability of the platform—for everyone. Dive in to learn more: https://lnkd.in/e-3Uutic

2

As Salesforce completes its acquisition of Informatica, it's worth a look back at my perspective from June on the then-proposed combination. "Data is integral to AI: large volumes of data are required to train models, while data freshness is important for inferencing in interactive applications and data quality is fundamental to ensuring that the output of agentic and generative AI (GenAI) initiatives can be relied upon. Poor data management can therefore be an impediment to AI. While AI-ready data is clean, well-organized and compliant with regulatory standards, too many enterprises find themselves struggling with data that is fragmented, inconsistent and not easily accessible." https://lnkd.in/e79A2djx

13

The right tools in the hands of a craftsman can result in amazing things. In neuroscience, our "tools" are the architectures that handle complex, multimodal brain data. When we treat data engineering as a craft - focusing on precision and scalability - we move from storing bits to enabling breakthroughs.

2

It's fascinating to watch how AI is reshaping org design in real time. The promise we've talked about for years — pushing decisions down to the people who are close enough to understand the problem — finally seems to be materializing. We're all watching this take shape most clearly in product development, especially at software and services companies where AI-driven delivery is the revenue engine. It's less obvious how this translates for more traditional companies, where IT and Centers of Excellence have been deliberately built as support functions to "the business." Do we consolidate them under the CFO, CRO and COO by business function? Or change even more radically with smaller, well-defined process and data domains? Either way, reconciling matrixed, work-oriented, nimble teams with the leaders who own financial and risk accountability will stay messy for a while.

2

1 Comment

Fantastic blog post from our friends at Whatnot on how they are unlocking trusted AI-powered analytics. Of course I am biased, because a portion of their solution is using Snowflake's Semantic Views. https://lnkd.in/gGNYugfZ A few highlights: 📈 How to navigate the trust/flexibility tradeoff without creating a mess (focus on evolving and enriching context to "hill-climb" to expanded analytic reach); 💡 How Semantic Views can use Cortex Search Services to do "smart matching" of terms in user queries (I don't know what a labubu is and I will probably spell it wrong anyway); 💻 AI doesn't make high quality data modeling and data development practices go away. Integrating the Semantic View lifecycle into existing dbt models and development best practices is key.

61

6 Comments