Osman Khadri Syed
Riyadh, Saudi Arabia
BI · Data Engineering · Applied AI
From KPI definition to semantic model to the dashboard a leadership team reviews each month, I connect business intelligence, data engineering and AI-assisted workflows. Most of my delivery has been in supply chain analytics: procurement, suppliers and inventory.
What I do
A KPI on a leadership dashboard is only as good as its definition, its model and the data under both. I work across all three, and I use AI tools where they speed up the work without weakening that chain.
Run requirements sessions with data owners, agree what each KPI means, then build the Power BI product and take it through stakeholder validation and UAT.
Procurement, SRM and Inventory dashboards; monthly Management Committee report moved from slides to Power BI.
Semantic models in SSAS Tabular, transformation logic in Power Query M, automated extraction from ERP reports, and layered pipelines with explicit quality gates.
30-table SSAS Tabular refactor; Bronze → Silver → Gold pipeline with 17 data-quality gates.
Specialised agents on Claude Code that share one body of project context, used to draft models, code and reviews. I keep design decisions, validation and anything that goes out the door.
Six-role agent workspace in daily use. Experimental, personal, not an enterprise deployment.
definition → model → data → product → validation
Selected work
Professional delivery, a personal engineering project, an AI workflow experiment and a model refactor. Each one states what was built, what was measured and what was not.
A multi-stream BI product for the Supply Chain Analytics Centre of Excellence: three workstreams, one monthly leadership reporting cycle, and data arriving from ERP reports, spreadsheets and manual exchanges.
Procurement, supplier and inventory performance lived in separate extracts, and monthly leadership reporting was assembled in slides.
KPI requirements agreed with data owners, modelled in SSAS Tabular and Power Query M, with Oracle EBS report extraction automated in Power Automate Desktop.
Procurement dashboard taken to go-live with UAT sign-off; SRM dashboard with 27 validated KPIs and a 52-KPI Vendor Profile Card; 15-KPI Inventory dashboard.
One governed place for Procurement, SRM and Inventory performance; the Management Committee report now runs in Power BI instead of PowerPoint.
Illustrative recreation. Table names are generic; no employer data, vendor names or screenshots are shown.
A delivery-monitoring model built around one governance question: when a giga-project re-baselines, is it still on track? Vision 2030 is the subject domain. This is not a government system and has no institutional affiliation.
Headline progress figures hide slippage when a plan is re-baselined, and project-level delivery data is messy: duplicates, stale versions, mixed date formats, mismatched names.
Bronze lands everything as text with lineage; Silver types, de-duplicates and logs every fix; Gold publishes a star schema. 17 quality gates fail the run on bad data.
Power BI project (PBIP/TMDL) with 30+ documented DAX measures, including schedule performance against both the original and the current baseline.
A PMO view that separates “on track against today's plan” from “on track against the original plan”, built on data that can be traced row by row.
Same project, two answers. Measuring against a re-written baseline makes slippage disappear, so the model reports both side by side. Values come from the project's synthetic progress series.
Counts are from the project's verification run: Silver output matched a hidden ground-truth file on all 540 rows.
is_synthetic throughout.Six role-based agents on Claude Code, each with its own written brief, all reading the same project files: models, KPI lists, status notes. One shared context, six specialist roles, instead of six copies of the data.
Modelling, pipeline, visual and automation questions each need a different kind of expertise, and a generic assistant forgets the project's constraints between sessions.
Written briefs per role, a shared folder as the single source of context, and explicit rules: check existing tables and measures before proposing new ones; state the grain of any new table.
On-demand reviews (“Data Model Check”, “Viz Check”, “Build Check” for M, DAX or automation drafts) plus one consolidated daily digest across all six roles.
Faster first drafts and reviews that respect real constraints. I haven't timed the effect yet, so there is no productivity figure.
Five of the six agent roles are shown. The sixth covers personal visibility and is left out here. Diagram is a structural summary.
A procurement-stream SSAS Tabular model had grown from Excel-based logic. I refactored and documented it so changes stop breaking reports, then built Power Automate flows that act on what the model shows.
Capability map
No percentages or star ratings. Each item says whether it comes from professional delivery, a personal project, exploration or an exam.
Decision-support products for operations and leadership.
Where KPI definitions become reusable logic.
Getting data to a state worth modelling.
Taking repetitive steps out of the reporting cycle.
The work that decides whether a dashboard gets used.
Under the dashboard
This is the pattern I design toward. Select a stage to see what I've implemented in professional work, what in a personal project, and what is still a reference design.
The AI + BI direction
A natural-language request is the easy part. What makes it safe is the governed model it runs against and the validation it passes before anyone sees the result. Proposed architecture
“Which suppliers are slipping on delivery this quarter, by category?”
The request maps to defined measures and dimensions, never to raw tables.
Modelling, transformation and visual agents propose changes and components from a shared template library.
Data-quality tests, reconciliation to known totals, KPI-definition checks, then human review.
Row-level security and workspace permissions apply to generated content exactly as they do to hand-built reports.
By people, up front: grain, keys and relationships are agreed and documented before any agent works against the model. Agents read the model; they don't invent it.
KPIs live as named measures with descriptions. Generated content can only reference existing measures, and changes to definitions go through a change-request approval, as in the KPI governance flow I built.
The same way as pipeline output: automated checks against known totals and domain rules, like the 17 gates in the Vision 2030 pipeline, plus a named human reviewer before publication.
Answers come through the semantic model, so existing row-level security and permissions still apply. No agent gets broader data access than the person asking.
A template fixes structure, measures and accessibility rules; the agent adjusts filters, focus and wording for the audience. The underlying definitions don't change.
By reconciling against a verified answer set and known totals, by logging which measures each answer used, and by treating unverifiable answers as drafts, not results.
Evidence
These numbers describe what was delivered and how large it was. They are not business outcomes, and they aren't presented as savings or adoption.
Professional story
I started by building the reports. The work that interests me now is underneath them: how a KPI is defined, how the model holds it, how reliable the data is, and where AI can safely take on part of the effort.
My recent work sits in supply chain analytics: procurement, supplier relationship management and inventory. I've taken dashboards through requirements, modelling, UAT and go-live, and I've refactored the semantic model behind them. In personal projects I go further into the engineering side, with layered pipelines, quality gates and AI-assisted workflows.
The problems I want to work on next are ones where reporting, data engineering and AI meet: governed semantic models that both dashboards and AI agents can use, reusable analytical products, and automation that removes manual steps from the reporting cycle.
Next step
I'm happy to talk through semantic models, data pipelines, KPI governance, or how AI fits into an analytics team.