Power BI tooling, model layer first
BIMKit takes the schema you already have and gives you everything that's missing —
a fully-explored model, realistic synthetic data, an AI that designs report pages,
and a clean .pbip you can open in Power BI Desktop.
Free during private beta · Your production data never leaves your machine
.bim fixtures validated
/ the gap
That's the position every Power BI developer ends up in eventually. The model is yours, but the data either doesn't exist yet, can't leave a secure environment, or is six tickets deep in someone else's queue. Every workaround is bad.
BIMKit is the layer that wasn't there. The schema goes in, a complete working report comes out — same model, same measures, same relationships. Just data you can actually use.
/ how it works
Each step takes seconds. Nothing leaves your machine unless you choose to send it.
Drag a .bim or .tmdl file in. BIMKit parses every detail — tables, columns,
measures, hierarchies, partitions, M expressions, DAX. It validates the model and
warns about issues at parse time.
BIMKit builds an interactive ERD in seconds. Layered or force-directed layout, hover highlighting on relationships, hierarchy detection. Read your own model again — even the ones you inherited.
Pick row counts (smart defaults are usually right). BIMKit generates synthetic data that's actually realistic — every foreign key resolves, every hierarchy is coherent, every column type is honoured.
Open the AI Report Designer. Tell it what you want: "executive overview,"
"trend by department," "add a margin measure." It writes the
visuals onto a live canvas. Export to .pbip and open in Power BI
Desktop. Done.
.pbip (TMDL + PBIR) — git-friendly text
/ what's inside
Six capabilities. Each one solves something specific you'd hit on real models.
/ the foundation
.bim parser in Python.17 column properties. 14 table properties. Hierarchies, partitions, M expressions, DAX measures with format strings, calculation groups, RLS roles. It even parses DATATABLE literals so calculated lookup tables come through with their data intact — no DAX engine required.
Foreign keys resolve. Hierarchies hold together — Aberdeen lives in Scotland, not South West. Calendar tables get real Year/Quarter/Month relationships. Bridge tables model power-law cardinality so visuals don't look artificially flat.
Chat with your model. The AI knows every measure, dimension, format string, and display folder. Ask for visuals, redesigns, new measures — it emits structured actions onto a live canvas. Provider-agnostic: Ollama, OpenAI, Claude.
Validate every measure against the data you just generated. DuckDB for fast, cross-platform DAX-style evaluation. Power BI Desktop via a managed Analysis Services subprocess for full-fidelity semantics. Auto-connects when Desktop is detected.
.pbip.
Skip the binary .pbix world. BIMKit emits a 100% text-based
.pbip project — TMDL for the model, PBIR for the report layout.
Diffs cleanly in git. Opens directly in modern Power BI Desktop.
14 visual types — card, KPI, bar, column, line, area, combo, pie, donut, treemap, gauge, waterfall, scatter, table, matrix. The working set, not the toy set. All driven from generated data. All configurable on the canvas.
/ under the hood
A few of the things that took the longest to get right.
Source detection (CSV, SQL, web, OLE DB). Date-range extraction from
List.Dates and #date(). Implicit FK lineage discovered from
Table.NestedJoin chains. Column rename tracking through
Table.RenameColumns.
Calculated lookup tables defined entirely as DAX DATATABLE()
expressions are parsed and the literal data preserved exactly — no DAX engine
required.
Geography of Country → Region → City uses a real lookup tree so child levels are valid descendants of their parent. Calendar of Year → Quarter → Month is derived from the date column on the same row.
Three profiles for many-to-many bridges: uniform, power-law (Zipf — the long tail you see in production), and clustered (models repeat-customer patterns).
The data dictionary scales with model size. Under 50 measures: full DAX, all columns. 50–200: compact form. 200+: minimal — names and format tags only. A 200-measure model fits where a naïve prompt would overflow.
The static prompt prefix (role, model context, dictionary) is marked with an
Anthropic cache_control breakpoint. The first turn pays a small
cache-write surcharge; subsequent turns within five minutes read at ~10% of
normal input rate.
/ editions
Both editions run the same core. The only difference is whether Power BI Desktop integration is on the table.
Windows desktop · single-binary portable app
Web-deployed · use from any browser
.pbip export — same output, downloadable as a zip
Both read the same .bim, generate the same data, design the same reports,
export the same .pbip. Use either, or both, depending on the day.
Free during private beta. Pro tier coming with the public launch — drop a message for an invite.
/ questions
.bim (TMSL) — the JSON serialisation of a tabular
model, exported by Tabular Editor 2 and emitted indirectly by Power BI Desktop
via .pbit/.pbip. Still the most common format in the wild.
.tmdl (TMDL) — the modern human-readable, indent-based
successor to TMSL. GA since August 2024 and the default in new Power BI Project
(PBIP) folders. BIMKit reads single-document TMDL exports directly.
Drop either format on the upload zone — BIMKit sniffs the content and dispatches automatically.
No. The parser is read-only. Generated data and any AI-created measures live in
the BIMKit canvas and the exported .pbip — never written back to the
source .bim.
Local edition: your real data never enters BIMKit, since BIMKit
only reads the .bim structure.
Hosted edition: your .bim is parsed in memory on the
server and discarded. Generated synthetic data and AI chat history live in your
browser session, not on our servers.
Tested clean against AdventureWorks 2 (14 tables, 22 relationships, 58 measures, 203 columns) and a custom 33-table production fixture. Adaptive prompt sizing keeps the AI workable at scale; the parser is unbounded.
Ollama (local, free, default for desktop), OpenAI (GPT-4o family), and Anthropic (Claude Haiku, Sonnet, Opus). Provider-agnostic protocol — adding a new one is a single Python file. Hosted edition defaults to Claude Haiku for cost efficiency.
Every AI-generated measure is run through real DAX evaluation against the generated data — DuckDB on hosted, DuckDB or Power BI Desktop locally. Errors are surfaced explicitly, not silently skipped.
Only the Power BI Desktop integration does — that's a Windows-only Microsoft product. The rest of BIMKit (parser, generator, DuckDB DAX engine, AI, PBIP export) runs anywhere. Hosted is the cross-platform path.
Currently a private project in beta. Get in touch below if you'd like access, want to talk roadmap, or have a specific use case in mind.
/ get in touch
Send a message — I'll get back to you. Particularly interested in models that break something interesting.