Notes from the operating loop. Built by data, for operators.
Practical writing on autonomous data, governed metrics, and the workflow that replaces the modern stack. No fluff. No SEO bait.
Latest field notes
22 results236 of 239 models built. The pipeline called it a failure.
One agent-generated dbt model referenced a dropped column, and a run that materialized 236 tables got collapsed into a single boolean: failed. Here's why partial materialization has to be a typed outcome with model names attached — and what an agent has to check before it answers a question about a half-rebuilt warehouse.
Structured output guaranteed the shape. It didn't guarantee the company existed.
Constrained decoding gives you a perfectly formed partner record every time — right fields, right types, nothing to parse. It says nothing about whether the company on that record was ever on the page. Here's the field-level evidence gate we put behind LLM extraction, and the boundary bug that let a half-invented company name through anyway.
The user closed the tab. The query didn't get the memo.
A browser tab closes mid-request and the backend never finds out — a tool call keeps running, a proxy keeps streaming, a warehouse connection stays open for a client that's gone. Here's why disconnect has to propagate as a real cancellation signal through every hop, not just get swallowed at the edge.
The dbt model that referenced nothing, and the run that kept going anyway.
An AI-generated dbt model refs a staging table that was renamed last week. dbt won't compile the project until that's fixed — which means every model in the run fails, not just the broken one. Here's why quarantining the orphan, not the run, is the only fix that scales.
The number was right. Nobody could tell you why.
An AI data agent can land on the correct number for the wrong reason — a join that fans out and a filter that happens to collapse it back down, this month. Here's why every autonomous answer needs a provenance trail, and what that trail actually has to contain.
Not every model gets to write the query that answers your revenue question
New frontier models ship every few weeks, and it's tempting to route every question to whichever one is newest. Here's why DataAgents gates model access behind a capability registry — certifying which LLM is trusted for which class of data question, and validating every answer regardless.
Schema drift is the outage nobody pages you for.
An upstream column gets renamed on a Tuesday. Nothing errors. Your pipeline keeps running, your dashboard keeps rendering, and your numbers are quietly wrong for three weeks. Here's why schema drift is the most expensive failure in data — and why an agent that tracks schema evolution is the only thing that catches it in time.
The Boring Parts Will Kill Your Startup Before Your Product Does
Your MVP is perfect. Your onboarding flow is slick. Your product-market fit is real. And your data operations are a lawsuit waiting to happen.
Vibe-Coding Your Data Pipeline Is Going to Hurt
Teams are delegating data pipeline architecture to AI agents with vague prompts. The result: fragile pipelines with silent failures, no lineage, and no one who understands how anything works. Here is why context-aware data agents are the only safe way to build autonomous analytics.
Your Data Platform Migration Will Cost 4x More Than Promised. Here’s Why.
A real team’s Databricks bill hit 2x their buffered estimate before they even finished 30% of the migration. Here’s why the modern data stack migration treadmill keeps producing expensive surprises — and what actually breaks at each step.
The autonomous data team is a workflow, not a team.
Why hiring three engineers, an analyst, and a BI tool no longer makes the business smarter and what replaces them when the platform owns the loop end-to-end.
A semantic layer that survives the next BI tool.
How we model business definitions in a way that stays correct when Slack, the web app, the API, and finance all ask the same question.
Why your contribution margin disagrees with your dashboards.
Six places refunds, shipping zones, and discount stacks quietly destroy DTC margin and how to model the truth in one definition.
The reporting backlog is a symptom. Here is the disease.
Why centralised data teams keep sliding into ticket queues, and what a self-serve operating loop looks like when it actually works.
Replacing 4 dashboards and a data engineer at Folie Studio.
How a fast-growing skincare brand collapsed Shopify, Meta, Klaviyo, and Stripe into one governed thread, and what they cut from the stack.
Backfills should be boring.
A note on idempotent ingestion, partition-aware retries, and why we let operators rerun a sync window from a Slack reply.
Deterministic agents over stochastic ones.
Why an answer the operator can defend matters more than an answer the model finds clever and how we draw that line in production.
The Future of Product Analytics Is Autonomous
BI tools gave teams dashboards. AI is about to give them answers. Here's what product analytics looks like when the machine understands your business.
5 Data Signals That Predict Feature Adoption Before It Happens
Most product teams measure feature adoption after the fact. Here are the leading indicators that let you intervene while there's still time.
How to Build a Data-Driven Product Roadmap That Stakeholders Actually Trust
Data-driven roadmaps fail when the data doesn't match stakeholder intuition. Here's how to build one that earns trust instead of sparking arguments.
The ROI of Predictive Analytics: What Finance Leaders Actually Want to Know
CFOs aren't opposed to predictive analytics - they're skeptical of vague ROI claims. Here's how to make a number-backed case.
Why Most New Features Fail - And What the Data Shows Beforehand
80% of new features don't reach meaningful adoption. The data signals are almost always there weeks before anyone acts on them.
See Your Data Clearly - Without Building a Data Team.
Connect your sources, standardize your metrics, and get decision-ready answers in minutes.