Case 01
25+ business units, one definition of margin
Multi-entity manufacturing group · three continents
Outcome at a glance
Before · a margin per entity
After · one governed measure
Sites and business units supported · 2017 to present
Business Intelligence · Financial Systems · Virginia, United States
I build the layer the numbersthe numbers come from.
Ten years turning factory floors, ledgers and disconnected systems of record into one set of figures that finance, operations and the board of a $2.2B+ manufacturing group can all argue from, and leading the team that produces them. Seven systems of record, 25+ business units, one definition of the truth.
Virginia, United StatesUS permanent resident
Keep going
A live reference architecture · four working demos · five problems worth the trouble
10
Years in BI and data, across manufacturing in the US, Middle East and Europe
25+
Business units reporting through one governed model, across three continents
13
Production dashboards maintained in parallel, across 29 concurrent projects
2.45m
Platform sessions since 2022, growing 24% year over year
01 — The architecture
Seven systems in, one number out
Every organisation I have worked in starts the same way: systems of record that do not speak to each other, and a finance team reconciling them by hand. This is the shape of the thing that fixes it — and the shape I build toward.
reference architecture · running
0rows/s
0models
0tests passed
0quarantined
02 — What I actually do
Three jobs that only look like one
Most BI roles are one of these. The work I have spent ten years on is all three at once, which is why the figure on a slide and the figure in the ledger match.
7 systems of record+ OPC telemetry and live market feeds
ERP, accounting, plant-floor telemetry, logistics and market prices, each on its own grain and calendar. Loads that run unattended and say so when they have not.
- Dynamics 365
- Oracle and Oracle JDE
- OPC telemetry
- REST APIs
- Scheduled loads
- Python
1 definitionper measure, reconciled to the ledger
The governed layer: margin is defined once, tested every run, and every report inherits it. The unglamorous half, and the half that decides whether two people in a meeting reconcile or just disagree louder.
- Dimensional modelling
- Semantic layer
- dbt
- Data quality tests
- Lineage
- Row-level security
2.45M sessionssince 2022, 13 production dashboards
Executive dashboards, close packs, forecasts and a natural-language agent over the same certified figures, so asking a question does not require knowing which table it lives in.
- Qlik Sense
- Power BI / DAX
- Tableau
- MCP + LLM
- Forecasting
03 — Live demos
Four things, actually running
Every one of these is real code computing over a fabricated dataset built for this site: six business units, five product families, thirty-two months. Nothing here comes from any employer. Click, filter, ask it something — the numbers recompute in the browser.
Revenue and gross margin by month
Bars are revenue. The line is gross margin percent.
Revenue
Gross margin %
12-month average
Mix by product family
Share of revenue and the margin each family carries.
Business unit performance
Click a column heading to sort. Click a unit to filter the charts above.
| Business unit | Revenue | GM % | YoY | Volume (t) | Share |
|---|
Built with no charting library — the SVG is generated from the data at render time. Filters, sorting and the aggregation all recompute client side, the same way a governed semantic layer recomputes a measure when you change a selection.
What this is
A natural-language layer over the same governed dataset the dashboard uses. It resolves a question to a measure, a dimension and a filter, runs the aggregation, and shows you the plan it used.
Why the plan matters
An agent over finance data is only useful if the answer is auditable. Every response here opens to reveal exactly what was computed — no figure appears that you cannot trace.
Try asking
- Which unit has the best margin?
- Revenue last quarter
- How did Northpoint do this year?
- Worst month for margin
- Compare Cascade and Ellery
Intent resolution here is deterministic pattern matching, not a language model call — a live model would need a key, and a portfolio page should not carry one. In production the same shape holds: the model picks the measure, dimension and filter, and a governed query engine does the arithmetic. That separation is the point. The model should never be the thing doing the maths.
Scheduled loads
Refreshing continuously. Watch a run start, progress and land.
| Pipeline | Source | Schedule | Rows | Freshness | Status |
|---|
Event log
Newest first.
Load duration, last 40 runs
The band is the normal range. Points outside it are what you want an alert for, not an average.
The thing that makes reporting trustworthy is not the dashboard, it is knowing the dashboard is built on data that landed. Freshness, row-count drift and duration outliers catch the three ways a pipeline lies to you: it did not run, it ran on nothing, or it ran on something new.
Revenue forecast
Holt linear trend with additive seasonality, fitted in the browser. Bands are 80 and 95 percent.
Actual
Forecast
80% interval
95% interval
Backtested by holding out the final six months, fitting on the rest and scoring the prediction — which is the only honest way to quote an accuracy figure. A forecast without a backtest is a drawing.
04 — Case studies
Five problems worth the trouble
Written at the level of the problem and the approach. Client figures, architecture and screenshots stay where they belong, which is with the employer who paid for them. Each one opens into a full write-up — the constraint that ruled out the obvious fix, the decisions I made, and what each decision cost.
Case 02
From the plant floor to the P&L
OPC line telemetry · ERP · conforming layer
Outcome at a glance
Before · plant clock vs ledger clock
After · one reconciled view
Case 03
A commercial programme nobody could see
Customer rebate lifecycle · end to end ownership
Outcome at a glance
Before · surprise at settlement
After · watched as it accrues, handed over in six weeks
Case 04
Asking a finance question in English
Natural-language layer over governed data
Outcome at a glance
Before · every question queued for a person
After · answered from certified figures, plan attached
Case 05
Journal entries that post themselves
Accounting automation · anomaly detection · reconciliation matching
Outcome at a glance
Before · every entry prepared by hand
After · routine entries post, exceptions go to review
05 — Stack
What I work in
Listed by how much of it I have actually run in production, not by how it looks on a keyword scan. Everything here is in production use; each tag says how.
Analytics platforms
- Qlik Sense and QlikView10 yrs
- Power BI, advanced DAXadvanced
- Tableauadvanced
- Qlik Cloud migrationled
- Semantic modellingcore
Data engineering
- Pythondaily
- SQL and T-SQLdaily
- dbtcore
- Scheduled loads and freshness monitoringautomated
- Dimensional and snowflake schema designdesigned
- Git, CI/CD, code reviewworking
Governance and quality
- Data quality testsautomated
- Reconciliation and validation controlsowned
- Anomaly detectionlive
- Fuzzy and semantic matchingbuilt
- Row-level securityenforced
Source systems
- Dynamics AX 2012, D365deep
- Oracle, Oracle Fusion, Oracle JDEintegrated
- Sageintegrated
- OPC line telemetrylive feed
- Quickbase, FoxPro, SharePointconnected
Cloud and applied AI
- Azure and Microsoft cloudhands-on
- AWS (S3)hands-on
- MCP and LLM toolingbuilt
- Prompt libraries and LLM evaluationowned
- Forecasting: ARIMA, SARIMA, Monte Carloapplied
- Linux and shellworking
Career
2022 — PRESENT
Senior Business Intelligence and Financial Systems Specialist
INDEVCO North America · United States2017 — 2022
Business Intelligence Specialist
NAPCO National · EMEA2015 — 2017
Project Manager
Born Interactive · Lebanon2013 — 2015
Software Engineer
BMW · Lebanon
Education
2025
M.S. Data Analytics
Western Governors University2018
B.S. Computer Science
Notre Dame University, LebanonProfessional development
03/2026
Financial Analysis and Business Projections
Continuing professional education registered with NASBA2024 – 2026
Project management coursework
Approved by the Project Management Institute (PMI)2026
Additional coursework
Microsoft, Google and Oracle technologies
Get in touch
Ten years of this work has mostly been a long conversation with people solving the same problems — reconciling systems that disagree, closing the books faster, making a number one that people trust. Always glad to have that conversation.
Email is the surest way to reach me.
- Emailgeorgio.hb@icloud.com
- Phone771-333-7771
- LinkedInin/georgio-hajj-boutros
- Based inVirginia, United States
ResumePDF, two pages. Current as of September 2026.
Download PDFAll data on this site is synthetic