Khuyen Tran

Glean is talking OLAP and Foundation Capital is talking OLTP. Database people know they both have a place. But there is something more interesting. > that forces a probabilistic model to deliver a deterministic outcome. The Ralph Wiggum solution to inject deterministic outcome in a probabilistic env? There is another way to do it: neuro-symbolic. It's probably the second most avoided word in our lingo after graph databases. Essentially RLM takes a neuro symbolic approach to generate code to do something and executing it instead of doing so using a probabilistic model. Now, apply the same solution to data. Probabilistic index is ok. But we can't make the primary data used in the decision making path to be probabilistic.

8

4 Comments

Coming closer to the heart of FinePhrase, the prompts. We tried 9 different prompts rephrasing the seed documents into polished synthetic data. Each prompt provides a different perspective. Four prompts outperform the DCLM baseline: Math, Table, FAQ and Tutorial. This experiment really drives the point home about how important the prompts are. The difference between the worst (commentary) and the best (math) is 0.3 points or 25%! Tomorrow it's all about model sizes.

18

Semantic Layers, Ontologies... I don't think anyone knows what the final version looks like, but we're getting closer. I've publicly benchmarked tens of different formats and contents and I want to do next level run at testing what context truly matters. You've all seen things that work for your agent, and I want to hear them, including your wild ideas. Here are some thoughts to spark the discussion: - Document the source systems, so that the agent knows what it is working with. - Share your company and department level goals with the agent, it will make it more opinionated and steer its answers in the right direction. - Standard dates to aggregate your metrics with. Let me know your thoughts, and I'll include them in my next writeup.

5

6 Comments

Every AI agent project I've seen stall had the same root cause: It wasn't the model, data, or vendor. It was the moment someone said, "We need a human to approve that," and nobody pushed back. Not because it was wrong, but because nobody had done the work to define what the agent could and couldn't do on its own. So the team bolted a human approval step onto every action and what started as an autonomous system became a recommendation engine with a nicer interface. The technology isn't the bottleneck on most AI projects. The org chart is. #AI #AIStrategy #Leadership #TechnologyLeadership

9

3 Comments

🚀 Agentic AI is changing how we build ML systems We’re moving from single-prompt LLM calls to stateful, multi-step agents that plan, reason, retrieve, and act. Key shifts I’m seeing in production: • DAG-based agent workflows (LangGraph > chains) • Tool-calling as first-class citizens • Memory + evaluation loops baked into agents Agentic systems aren’t “prompt engineering 2.0” — they’re distributed systems with reasoning.

1

Here I’m sharing a few chapters from a practical book I’m currently writing on LLM fine-tuning for industry practitioners. It covers fine-tuning data preparation, data templates, model selection, fine-tuning strategies, and efficient deployment—all from a hands-on, production-oriented perspective. The goal is to provide a practical toolbook for real-world LLM applications. 🚀

15

Our new RAG app was confidently wrong 60% of the time. We were building a Q&A bot over our internal technical documentation. The LLM was a top-tier model, so we figured it would be easy. But the answers were terrible. It would hallucinate features that didn't exist or pull completely irrelevant paragraphs. It looked like a classic model failure. The problem wasn't the model. It was the "R" in RAG. Our naive chunking strategy was splitting sentences and paragraphs right down the middle, destroying the context before the retriever even did its job. The retriever was grabbing nonsense snippets, and the LLM was just dutifully summarizing the nonsense we fed it. We spent two weeks re-building our chunking and retrieval logic to be context-aware. The failure rate dropped to under 10%. Don't blame the generator for a bad retrieval. Your RAG system is only as smart as the context you find for it. #MachineLearning #RAG #DataEngineering

10

Free the credits your AI roadmap needs on #Snowflake. Dynamic tables refresh on whatever target lag you gave them, whether or not the source actually changed. Set the lag to a minute on a table that's written once a day and you're paying for the schedule, not the data. The refreshes land as system-generated work with no user attached, so nobody goes looking. Anavsan walks those refresh credits back to the lag setting itself, and to the DDL that put it there. Then you try the looser lag in a sandbox against a week of real source activity and watch staleness before anything ships. When the refreshes stop, it shows up in billed credits. Not a forecast. Most of these lags got copied from an example. They didn't get revisited. https://lnkd.in/ewYNgWV5 #Snowflake #FinOps #DataEngineering #DataOps #Shorts

5

Hot take for anyone early in their career 👀 Yes, seniors have years of experience. Context. Battle scars. That still matters. But with AI tools? Everyone is new. There is no “10 years of experience” with agents, copilots, or AI-first workflows. We’re all testing, breaking things, guessing, learning in public. And that’s where juniors actually have an edge right now. You’re: • Less attached to “the way we’ve always done it” • More willing to experiment • Faster to learn tools that don’t have best practices yet A lot of old career advice doesn’t apply anymore. “Just grind for 5 years” doesn’t help when the market is frozen. “Wait your turn” doesn’t work when the rules are changing monthly. But curiosity does. Shipping tiny experiments does. Sharing what you learn does. If you’re junior and feeling stuck: this is one of the rare moments where newness is an advantage. Lean into it.

28

3 Comments