Egor Tarasenko

data products vs semantic layers — not the same thing saw a good post from Tim Gasper on this and it stuck with me. his framing: data products are about the promise (who owns it, what quality to expect, who to call when it breaks), semantic layers are about the meaning.

so the real question isn't "should we build a semantic layer", it's "do we keep letting it live in 40 dashboards or do we make it define-once". we went the define-once route: metrics in dbt yaml -> auto-synced to Cube.js and BI -> even queryable back inside the dbt DAG.

data products vs semantic layers — not the same thing saw a good post from Tim Gasper on this and it stuck with me. his framing: data products are about the promise (who owns it, what quality to expect, who to call when it breaks), semantic layers are about the meaning.

A few months ago I started Broken Pipelines -- a Telegram channel about data engineering. Yesterday it passed 100 subscribers 🎉 Quick story of why it exists.

So I started writing: honest takes on new tools (no, DuckLake didn't kill Iceberg), production war stories, what actually breaks at 3am and why.

A few months ago I started Broken Pipelines -- a Telegram channel about data engineering. Yesterday it passed 100 subscribers 🎉 Quick story of why it exists.

I'm Egor, a Data & AI Engineer. I spend my days building data platforms -- Airflow pipelines, Trino, dbt, lakehouses, non-generic plugins, custom data solutions, and open-source projects. All the plumbing that turns raw chaos into data someone can actually trust.

"dbt Core vs Cloud" is a false binary. dbt is five decisions -- development, execution, CI, orchestration, metadata -- and you don't have to answer them the same way. Where I landed: development + execution in Cloud (the IDE won). CI in Core -- Cloud's CI is opinionated, my workflow disagrees.

Orchestration in Airflow: triggering Cloud jobs is one operator; wiring non-dbt dependencies around Cloud's scheduler is webhooks. And keep a dbt Core backup image regardless. same code, same repo, a DR plan that costs one Dockerfile. When dbt Cloud is down, "fully managed" means fully waiting.

OpenSpec sat in my feed for 4 months. I kept scrolling -- I've watched too many "spec-driven" frameworks become folders nobody opens. Then I had 3 people and an agent in one dbt repo and nobody could say what our models were even supposed to do. So I finally tried it.

Simpler than the hype. Two folders in git: specs/ is the truth, changes/ is work in flight like a feature branch. What won me over: spec review is separate from code review. A spec diff shows what behavior changed -- readable in seconds, not a 600-line PR at 2am.

We built an AI chat for dbt that understands your project as a knowledge graph. It was powerful. But it understood relationships, not meaning. So we fixed that 👇 newsletter.ponder.co/p/the-dbt-ai...

Fix #3: reranking via Cohere on Bedrock. KNN gets the ballpark right but the ordering wrong. A reranker reads query + candidate together in one pass -- cross-attention, not independent vectors. Run it on a shortlist of 20-50, not the whole dataset.

We built an AI chat for dbt that understands your project as a knowledge graph. It was powerful. But it understood relationships, not meaning. So we fixed that 👇 newsletter.ponder.co/p/the-dbt-ai...

Fix #1: embeddings stored directly on each FalkorDB node. No separate vector store. No parallel infrastructure to maintain. Upload your manifest and catalog -- every node is searchable by meaning before you open the chat.

We built an AI chat for dbt that understands your project as a knowledge graph. It was powerful. But it understood relationships, not meaning. So we fixed that 👇 newsletter.ponder.co/p/the-dbt-ai...

I was on holiday in the midwest when Fable 5 launched. Figured nothing crucial happens in a week. Turns out a week is too much.

And the chain: Anthropic's largest investor (Amazon, $13B in) bypassed the company to brief Treasury directly, and within 24 hours the product was offline under threat of criminal penalties. Same pattern as Huawei, TikTok, the 1990s encryption wars.

I was on holiday in the midwest when Fable 5 launched. Figured nothing crucial happens in a week. Turns out a week is too much.

The trigger: asking the model to read a codebase and fix bugs. The government called it a jailbreak. GPT-5.5 does the same thing with no bypass needed.

I was on holiday in the midwest when Fable 5 launched. Figured nothing crucial happens in a week. Turns out a week is too much.

semantic layers keep solving the definition problem. nobody solves the query interface problem. www.linkedin.com/posts/burakk...

define once they said. you now have 7 semantic layers, final_final_v2, and a prayer. a semantic layer without multiple query interfaces is just well-documented SQL. is owning the full stack the only way to actually solve the semantic layer problem?

I just received some photos from my talk with Alex Selkirk at SQLBits 2026, where we had two sessions about solutions we've built at Ponder Labs -- so I figured it's a great opportunity to recap what we covered, share the blog posts again, and show off my beautiful photo 😊

dbt projects become dependency webs nobody can navigate. We loaded the full dbt graph into a graph DB and put an AI chat on top. Ask "what's the most reused source?" and actually get the right answer. newsletter.ponder.co/p/graph-powe...

I just received some photos from my talk with Alex Selkirk at SQLBits 2026, where we had two sessions about solutions we've built at Ponder Labs -- so I figured it's a great opportunity to recap what we covered, share the blog posts again, and show off my beautiful photo 😊