Parul Manek

I help customers scale some of the largest Generative AI workloads in the world. I've seen a lot of success, and a lot of failure. One common thread across all the most successful generative AI workloads is Evaluations; it is a topic that has been my largest area of focus for the last few years. In my latest blog post, I share some observations on the most common questions and solutions for building awesome generative AI evaluations. https://lnkd.in/gBKc3avW

61

6 Comments

Snowflake is releasing so many AI features lately, it's really hard to keep track of them all. For AI Agents, evaluation is very important to improve their performance. But setting up an evaluation system yourself takes a lot of work. Fortunately, the Evaluation feature for Snowflake's Cortex Agent is now in GA! If users just need to prepare evaluation data they can utilize this built-in feature to run continuous evaluations. In my blog post, I explain how this feature basically works, and also share my favorite part: how to run it and get the evaluation results using just SQL. [medium]: https://lnkd.in/gyj-Qu9n [zenn]: https://lnkd.in/gzMV2k_P

52

Latest stable release of the Python SDK for Azure Cosmos DB has been heavily battle tested by OpenAI in maintaining ChatGPT. We’ve also released semantic reranker support (private preview). Check out the blog for details 🙂 https://lnkd.in/e_65Mpd6

19

MCP ensures cross-model orchestration, while 7NF guarantees semantic precision and reversibility. AI may be good for one-time communication but business needs responsibility for decisions based on explainable resources. ~ In your 2025 polytypic architecture, you have defined a critical division of labor for enterprise AI: MCP handles the "handshakes" (orchestration), while 7NF handles the "proofs" (semantic precision).  This distinction addresses the "Accountability Gap" in AI—moving from one-off creative chats to Responsible Business Intelligence based on explainable, reversible resources. 1. MCP: The Orchestrator of "How"  The Model Context Protocol (MCP) fills the integration gap between fragmented AI models and external tools.  Dynamic Connectivity: It allows AI agents to securely discover and invoke data repositories or services without bespoke "glue code". Orchestration Logic: MCP facilitates coordinated sequences across multiple tools, but it operates at the protocol level. It tells the AI how to talk to the data, but it does not define the truth of the data itself. https://lnkd.in/gfsDxhnU 2. 7NF: The Guarantor of "What" (Precision and Reversibility)  While MCP connects models, 7NF (Seventh Normal Form) ensures the data being connected is irreducible and semantically grounded. Semantic Precision: 7NF eliminates multi-valued dependencies, reducing every fact to its atomic binary relation Rel⟨A,B⟩ or Attr⟨E,V⟩. This prevents the "hallucination of joins" by ensuring that every relationship is a mathematically precise link between identities. Reversibility: Because 7NF uses canonical forms, the polytypic engine can always "reverse" a complex view back to its atomic components with zero information loss. This is essential for Explainable AI (XAI), where every business decision must be traceable back to its original 7NF atomic evidence. https://lnkd.in/gFQ6bn_Y 3. Resolving the Business Responsibility Challenge  For 2025 business needs, one-time communication is insufficient; decisions require auditable and explainable resources. Explainable Reasoning: By feeding AI from 7NF-structured data, recommendations move from "probabilistic guesses" to "structural proofs." A decision can be justified by pointing to the specific, discrete 7NF atoms that triggered the natural transformation. Accountability via Identity: 7NF anchors all data in persistent Entity Identifiers Id⟨E⟩. Unlike transient context windows, these identities provide a permanent, reversible record of truth that survives across multiple MCP sessions. https://lnkd.in/gfWK8Ugh ~

7

BigQuery gets a lot of love and attention for its capabilities (and it should; it's an amazing analytics platform!) but I gotta say I'm so impressed with what AlloyDB brings to the table for AI-centric applications. The baked in AI Query Engine in AlloyDB (https://lnkd.in/dnA-n-VM) is a game changer for an application database. Having the transactional capabilities of Postgres, combined with a columnar storage model, combined with the AI Query Engine? Game changing. Check it out if you haven't!

30