<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>Piotr Zawieja — Blog</title>
    <link>https://goldlayer.dev/writing/</link>
    <description>Analytics engineering, dbt, Snowflake, and AI-assisted data systems.</description>
    <language>en</language>
    <atom:link href="https://goldlayer.dev/writing/feed.xml" rel="self" type="application/rss+xml"/>
    <item>
      <title>Can you trust this answer? Evaluate the whole analytics agent</title>
      <link>https://goldlayer.dev/writing/analytics-agent-evaluation-judges.html</link>
      <guid isPermaLink="true">https://goldlayer.dev/writing/analytics-agent-evaluation-judges.html</guid>
      <pubDate>Tue, 17 Nov 2026 09:00:00 +0000</pubDate>
      <description>Build useful analytics-agent evaluations with deterministic checks, calibrated LLM judges, human review, and cases that test the whole answer path.</description>
    </item>
    <item>
      <title>Give your analytics agent the right context, not the whole warehouse</title>
      <link>https://goldlayer.dev/writing/analytics-agent-context-engineering.html</link>
      <guid isPermaLink="true">https://goldlayer.dev/writing/analytics-agent-context-engineering.html</guid>
      <pubDate>Mon, 16 Nov 2026 20:37:55 +0000</pubDate>
      <description>A practical way to assemble relevant, current context for analytics agents without dumping every metric, schema, and chat turn into the prompt.</description>
    </item>
    <item>
      <title>Give your analytics agent a contract, not a bigger prompt</title>
      <link>https://goldlayer.dev/writing/analytics-agent-semantic-contract.html</link>
      <guid isPermaLink="true">https://goldlayer.dev/writing/analytics-agent-semantic-contract.html</guid>
      <pubDate>Mon, 16 Nov 2026 14:00:00 +0000</pubDate>
      <description>A clear, practical guide to defining metric meaning, grain, time, joins, and access before an AI agent queries business data.</description>
    </item>
    <item>
      <title>Agents on your semantic layer, Part 2: Three additions that make it production-safe.</title>
      <link>https://goldlayer.dev/writing/semantic-layer-agent-harden.html</link>
      <guid isPermaLink="true">https://goldlayer.dev/writing/semantic-layer-agent-harden.html</guid>
      <pubDate>Tue, 30 Jun 2026 08:20:16 +0000</pubDate>
      <description>Add disambiguation pre-flight, template matching, and three-tier memory to your dbt Semantic Layer agent. Part 2 of 2 — the hardening layer.</description>
    </item>
    <item>
      <title>Agents on your semantic layer, Part 1: Three tools. One loop. A working analytics agent.</title>
      <link>https://goldlayer.dev/writing/semantic-layer-agent-build.html</link>
      <guid isPermaLink="true">https://goldlayer.dev/writing/semantic-layer-agent-build.html</guid>
      <pubDate>Tue, 30 Jun 2026 08:20:12 +0000</pubDate>
      <description>Build a real analytics agent on dbt Semantic Layer and Snowflake: three tools, one Claude loop, about 80 lines of Python. Part 1 of 2.</description>
    </item>
    <item>
      <title>Agents on your semantic layer, Part 0: Your analytics agent is confident. That's the problem.</title>
      <link>https://goldlayer.dev/writing/agents-on-semantic-layer-101.html</link>
      <guid isPermaLink="true">https://goldlayer.dev/writing/agents-on-semantic-layer-101.html</guid>
      <pubDate>Tue, 30 Jun 2026 08:20:08 +0000</pubDate>
      <description>Semantic layer agents fail silently. Here's what breaks, how to prevent hallucinations, and why your agent should ask before it answers.</description>
    </item>
    <item>
      <title>40 data alerts at 6am. AI triaged them in 90 seconds.</title>
      <link>https://goldlayer.dev/writing/ai-triage-data-monitoring-alerts.html</link>
      <guid isPermaLink="true">https://goldlayer.dev/writing/ai-triage-data-monitoring-alerts.html</guid>
      <pubDate>Sun, 28 Jun 2026 08:08:43 +0000</pubDate>
      <description>Data monitoring tools produce more alerts than humans can triage. Here's the classification framework and prompt structure to separate signal from noise.</description>
    </item>
    <item>
      <title>Six analytics jobs I've handed to AI — none of them are writing SQL</title>
      <link>https://goldlayer.dev/writing/ai-analytics-not-sql-generation.html</link>
      <guid isPermaLink="true">https://goldlayer.dev/writing/ai-analytics-not-sql-generation.html</guid>
      <pubDate>Sun, 28 Jun 2026 07:45:06 +0000</pubDate>
      <description>SQL generation demos well. Here's where AI actually saves the most time in analytics work — and none of it is writing code.</description>
    </item>
    <item>
      <title>You've been flattening Snowflake JSON wrong</title>
      <link>https://goldlayer.dev/writing/snowflake-variant-flatten.html</link>
      <guid isPermaLink="true">https://goldlayer.dev/writing/snowflake-variant-flatten.html</guid>
      <pubDate>Sat, 27 Jun 2026 08:18:53 +0000</pubDate>
      <description>LATERAL FLATTEN is the wrong default for Snowflake JSON. Dot-notation, parsed columns, and when to actually flatten — with SQL for each.</description>
    </item>
    <item>
      <title>How to know if your analytics agent still works: a deterministic eval harness</title>
      <link>https://goldlayer.dev/writing/analytics-agent-eval-harness.html</link>
      <guid isPermaLink="true">https://goldlayer.dev/writing/analytics-agent-eval-harness.html</guid>
      <pubDate>Fri, 26 Jun 2026 22:38:44 +0000</pubDate>
      <description>Scoring, cost tracking from real runs, and regression gating for a dbt analytics agent — without LLM-as-judge.</description>
    </item>
    <item>
      <title>Building an analytics agent on a real dbt project: what actually broke</title>
      <link>https://goldlayer.dev/writing/dbt-agentic-engine-build-log.html</link>
      <guid isPermaLink="true">https://goldlayer.dev/writing/dbt-agentic-engine-build-log.html</guid>
      <pubDate>Fri, 26 Jun 2026 22:38:39 +0000</pubDate>
      <description>A build log for dbt-agentic-engine — four tools, a hard call budget, a structured trace, and the things that only broke on live API calls.</description>
    </item>
    <item>
      <title>Revenue Intelligence Agent: an AI co-pilot that stops Customer Success from firefighting</title>
      <link>https://goldlayer.dev/writing/revenue-intelligence-agent-ai-cs-copilot.html</link>
      <guid isPermaLink="true">https://goldlayer.dev/writing/revenue-intelligence-agent-ai-cs-copilot.html</guid>
      <pubDate>Tue, 23 Jun 2026 12:12:52 +0000</pubDate>
      <description>A demo and build log for Revenue Intelligence Agent — an AI co-pilot that turns a renewal pipeline into daily, proactive CS action.</description>
    </item>
    <item>
      <title>Reconciling competing metric definitions across teams with an LLM</title>
      <link>https://goldlayer.dev/writing/reconciling-metric-definitions-llm.html</link>
      <guid isPermaLink="true">https://goldlayer.dev/writing/reconciling-metric-definitions-llm.html</guid>
      <pubDate>Sun, 21 Jun 2026 21:29:05 +0000</pubDate>
      <description>An LLM can't decide what 'active customer' means. It can find every definition you have, show where they diverge, and draft the doc.</description>
    </item>
    <item>
      <title>Forecasting with an LLM vs. a real time-series model: where each one is the right call</title>
      <link>https://goldlayer.dev/writing/forecasting-llm-vs-time-series-model.html</link>
      <guid isPermaLink="true">https://goldlayer.dev/writing/forecasting-llm-vs-time-series-model.html</guid>
      <pubDate>Sun, 21 Jun 2026 11:13:35 +0000</pubDate>
      <description>An LLM is good at explaining a forecast, not producing one. Where Prophet or ETS still beats asking Claude to predict next quarter's numbers.</description>
    </item>
    <item>
      <title>Snowflake clustering keys: what they are, when they help, and when to skip them</title>
      <link>https://goldlayer.dev/writing/snowflake-clustering-keys-guide.html</link>
      <guid isPermaLink="true">https://goldlayer.dev/writing/snowflake-clustering-keys-guide.html</guid>
      <pubDate>Tue, 16 Jun 2026 18:27:14 +0000</pubDate>
      <description>Clustering keys improve query performance by rearranging micro-partitions — but they cost credits continuously. Here is when the trade-off is worth it.</description>
    </item>
    <item>
      <title>Idempotency in dbt: the one pipeline property worth obsessing over</title>
      <link>https://goldlayer.dev/writing/dbt-idempotency-pipelines.html</link>
      <guid isPermaLink="true">https://goldlayer.dev/writing/dbt-idempotency-pipelines.html</guid>
      <pubDate>Tue, 16 Jun 2026 08:05:07 +0000</pubDate>
      <description>Most dbt models look idempotent but aren't. What breaks it silently and the concrete patterns — unique_key, surrogate keys, partition replacement — that fix it.</description>
    </item>
    <item>
      <title>AI for analytics engineering: where it actually helps and where it doesn't</title>
      <link>https://goldlayer.dev/writing/ai-for-analytics-engineering.html</link>
      <guid isPermaLink="true">https://goldlayer.dev/writing/ai-for-analytics-engineering.html</guid>
      <pubDate>Sun, 14 Jun 2026 12:31:51 +0000</pubDate>
      <description>Not a hype piece. Where AI tools earn their place in an analytics engineering workflow — and where they consistently fall short.</description>
    </item>
    <item>
      <title>dbt packages worth using in 2026</title>
      <link>https://goldlayer.dev/writing/dbt-packages-worth-using-2026.html</link>
      <guid isPermaLink="true">https://goldlayer.dev/writing/dbt-packages-worth-using-2026.html</guid>
      <pubDate>Sun, 14 Jun 2026 12:19:55 +0000</pubDate>
      <description>An honest assessment of which dbt packages actually earn their dependency weight — and which ones add complexity without proportional value.</description>
    </item>
    <item>
      <title>Dimensional modeling in dbt: the patterns that don't fit the textbook</title>
      <link>https://goldlayer.dev/writing/dimensional-modeling-dbt-practical-patterns.html</link>
      <guid isPermaLink="true">https://goldlayer.dev/writing/dimensional-modeling-dbt-practical-patterns.html</guid>
      <pubDate>Sun, 14 Jun 2026 12:19:51 +0000</pubDate>
      <description>Kimball tells you what to build, dbt gives you the tools. Where SCD Type 2 snapshots, bridge tables, and role-playing dimensions actually live.</description>
    </item>
    <item>
      <title>How to write dbt tests that catch real business logic errors (not just nulls)</title>
      <link>https://goldlayer.dev/writing/dbt-tests-business-logic-errors.html</link>
      <guid isPermaLink="true">https://goldlayer.dev/writing/dbt-tests-business-logic-errors.html</guid>
      <pubDate>Fri, 12 Jun 2026 13:00:00 +0000</pubDate>
      <description>Most dbt test suites catch infrastructure failures like nulls and duplicate keys, but miss the business logic errors that surface in weekly business reviews.</description>
    </item>
    <item>
      <title>dbt Semantic Layer vs. homegrown metric marts: when to use which</title>
      <link>https://goldlayer.dev/writing/dbt-semantic-layer-vs-metric-marts.html</link>
      <guid isPermaLink="true">https://goldlayer.dev/writing/dbt-semantic-layer-vs-metric-marts.html</guid>
      <pubDate>Fri, 12 Jun 2026 12:00:00 +0000</pubDate>
      <description>A practical decision framework for analytics engineers choosing between pre-aggregated metric marts and the dbt Semantic Layer with MetricFlow.</description>
    </item>
    <item>
      <title>Snowflake Cortex Analyst in practice: what the demo doesn't show you</title>
      <link>https://goldlayer.dev/writing/snowflake-cortex-analyst-in-practice.html</link>
      <guid isPermaLink="true">https://goldlayer.dev/writing/snowflake-cortex-analyst-in-practice.html</guid>
      <pubDate>Fri, 12 Jun 2026 11:00:00 +0000</pubDate>
      <description>Cortex Analyst looks smooth in a demo. Here is what actually happens when you point it at a real warehouse in week two.</description>
    </item>
    <item>
      <title>AI-assisted SQL review: using Claude to catch data quality issues before they ship</title>
      <link>https://goldlayer.dev/writing/ai-assisted-sql-review-claude-dbt.html</link>
      <guid isPermaLink="true">https://goldlayer.dev/writing/ai-assisted-sql-review-claude-dbt.html</guid>
      <pubDate>Fri, 12 Jun 2026 10:00:00 +0000</pubDate>
      <description>Not 'AI is transforming data teams' — a concrete workflow: the prompt, what it catches reliably, and where it fails. For analytics engineers who review dbt PRs.</description>
    </item>
    <item>
      <title>Late-arriving data: the silent failure that breaks your incremental models</title>
      <link>https://goldlayer.dev/writing/late-arriving-data-dbt-incremental.html</link>
      <guid isPermaLink="true">https://goldlayer.dev/writing/late-arriving-data-dbt-incremental.html</guid>
      <pubDate>Thu, 11 Jun 2026 11:00:00 +0000</pubDate>
      <description>Your dbt run succeeds. Tests pass. The dashboard looks fine. The data is still wrong. Late-arriving events and the dbt patterns that handle them correctly.</description>
    </item>
    <item>
      <title>dbt incremental models: the complete guide</title>
      <link>https://goldlayer.dev/writing/dbt-incremental-models-complete-guide.html</link>
      <guid isPermaLink="true">https://goldlayer.dev/writing/dbt-incremental-models-complete-guide.html</guid>
      <pubDate>Thu, 11 Jun 2026 10:00:00 +0000</pubDate>
      <description>How incremental models work, when each strategy applies, and the patterns that only surface after something breaks in production.</description>
    </item>
    <item>
      <title>How I set up Claude Code for analytics engineering work</title>
      <link>https://goldlayer.dev/writing/claude-code-analytics-setup.html</link>
      <guid isPermaLink="true">https://goldlayer.dev/writing/claude-code-analytics-setup.html</guid>
      <pubDate>Mon, 08 Jun 2026 10:00:00 +0000</pubDate>
      <description>Most analytics engineers re-explain their stack every AI session. CLAUDE.md, skills, and a few conventions make Claude useful from the first prompt.</description>
    </item>
    <item>
      <title>Why governed AI on your warehouse beats RAG on your docs</title>
      <link>https://goldlayer.dev/writing/governed-ai-warehouse-beats-rag.html</link>
      <guid isPermaLink="true">https://goldlayer.dev/writing/governed-ai-warehouse-beats-rag.html</guid>
      <pubDate>Mon, 08 Jun 2026 09:00:00 +0000</pubDate>
      <description>Most enterprise data AI projects start by indexing documentation — here is why that is backwards, and what a warehouse-native approach looks like instead.</description>
    </item>
    <item>
      <title>How to build a modern analytics data platform: the decisions nobody documents</title>
      <link>https://goldlayer.dev/writing/data-platform-decisions.html</link>
      <guid isPermaLink="true">https://goldlayer.dev/writing/data-platform-decisions.html</guid>
      <pubDate>Mon, 08 Jun 2026 08:00:00 +0000</pubDate>
      <description>Everyone writes about the tools. Nobody writes about the architectural choices made in the first three months that you cannot change without rebuilding.</description>
    </item>
    <item>
      <title>From Excel to dbt: a realistic transition guide for Polish data teams</title>
      <link>https://goldlayer.dev/writing/excel-to-dbt-polish-data-teams.html</link>
      <guid isPermaLink="true">https://goldlayer.dev/writing/excel-to-dbt-polish-data-teams.html</guid>
      <pubDate>Fri, 05 Jun 2026 11:00:00 +0000</pubDate>
      <description>A practitioner's guide to migrating analytics from spreadsheets to dbt, covering the governance and trust-building steps that most teams skip.</description>
    </item>
    <item>
      <title>Why your dbt project has 47 models and nobody trusts any of them</title>
      <link>https://goldlayer.dev/writing/dbt-47-models-nobody-trusts.html</link>
      <guid isPermaLink="true">https://goldlayer.dev/writing/dbt-47-models-nobody-trusts.html</guid>
      <pubDate>Fri, 05 Jun 2026 10:00:00 +0000</pubDate>
      <description>The dbt trust problem isn't a tooling issue — it's architectural. It shows up in every team that moved fast and skipped the governance layer.</description>
    </item>
  </channel>
</rss>