Karthik BI · Senior BI Engineer
Beyond the Dashboard
18 years turning data into decisions across manufacturing, finance, and tech. Here I share what I learn.
PBIP Gave Us Git. Skills Made Everyone Use It the Same Way.
Most of my team were new to Git. So we packaged the PBIP workflow — branch, change, commit, push, PR — as skills and scripts, with the ticket number carried through to promotion.
We Built a Tool to Find Unused Measures. Then the Platform Kept Shrinking It.
Hundreds of unused measures, found once a quarter by uploading files. Then PBIP, Git and a metadata database turned cleanup into a query.
A Year on PBIP: The Model Stopped Being a File
One year on Power BI Projects (PBIP): version control, multiple developers, pull requests from other teams and AI agents reading the model for context.
How I Built a Fabric App on 1,851,282 Flights
1,851,282 US flights, a semantic model and Claude Code. How I went from New item to a deployed Fabric App with a map, filters and animation.
Direct Lake Gave Me the Right Number. It Just Didn't Use Direct Lake.
A table based on a SQL view in a Direct Lake on SQL model returned the right total and still showed Direct Lake. Performance Analyzer showed it had gone through DirectQuery.
What Direct Lake Actually Is: I Added a Row and Never Refreshed the Model
I inserted one row into a Fabric Warehouse table and watched a Power BI report pick it up without a manual model refresh, and what that says about how Direct Lake works.
18 Fields in One Flat List, Half of Them Nobody Should Touch
Raw lat/long, an internal system flag, a vendor ID — all sitting in the field list next to the measures people actually need. Hiding and foldering it costs five minutes once.
Copilot Ignored a Real Field Twice. A Description Didn't Help. Synonyms Did.
'Extra' is an ambiguous field name. Copilot grabbed the wrong field, then grabbed another wrong one after I added a description. Synonyms were the only lever that worked.
The Total Row Said $111,270.85. The Real Answer Was $146,626.18.
A measure's total row doesn't sum what's on screen — it re-evaluates in its own filter context. For a table of daily maximums, that's a different number, and it's wrong for what it claims to be.
One Calculated Column Cost 1.4MB on a 10.9M-Row Table. The Measure Cost Nothing.
On sample data the difference is invisible. At 10.9 million rows, a single calculated column added over a megabyte to the file — and the identical logic as a measure added zero.
I Built the Same View-Switcher Two Ways: Bookmarks vs Field Parameters
Bookmarks won on setup speed. Field parameters won on object count, on what actually gets swapped, and on every view I added afterward.
One Checkbox, 55KB, and Only Two Date-Heavy Tables
Auto Date/Time builds a hidden Year/Quarter/Month/Day hierarchy for every date column whether you ask for it or not. On a two-table model that was 55KB — about 15%.
Sum of UnitPriceDiscount Showed $8.4 — A Number That Means Nothing
Drag a discount rate into a table and Power BI sums it by default. The total row read $8.4, formatted as currency, looking exactly like something you would trust.
A Text-Typed Date Column Doesn't Announce Itself as Broken
OrderDate came in as text. Nothing errored — the line chart just sorted Aug → Jul → Jun, and Relative date was missing from the filter pane entirely.
Auto-Detect Guessed My Cardinality, and Guessed It Wrong
Power BI's auto-detected relationship looked right for weeks — 1-to-1, bidirectional — until a customer placed a second order and refresh failed outright.