Flagship Independent Build

Nucleus

Executive Analytics & Decision Intelligence Platform

Most analytics estates fail the same way. Numbers live in fragmented reports, KPI definitions drift between teams, executives wait days for an analyst to run something, and forecasting or AI work sits disconnected from the governed data underneath. Nucleus is my answer to that problem, designed and built independently end to end.

Governed Metric LayerForecast + AnomaliesPlayer & Experiment AnalyticsRead-only AI Query Layer

What I built

One platform, owned end to end.

I designed and built Nucleus myself: the data architecture, the governed metric layer, the executive portal UX, the statistical engines, the access-control model, and the AI answer layer that sits on top of all of it.

The organising idea is a single governed metric layer. Formulas such as hold, RTP, rotation and net/store/day are defined centrally in SQL and governed views. Every consumer — a dashboard, a forecast, an anomaly score, or an AI answer — resolves the same definition. Nothing gets to invent its own version of a KPI.

On top of that layer, Nucleus adds boardroom-grade executive screens, Holt-Winters forecasting with confidence bands, severity-ranked anomaly detection, cohort and segmentation analytics, Taguchi-style experimental design for promotions, and a natural-language query layer restricted to read-only governed tools.

Architecture

Data flow, with governance as a cross-cutting layer.

01

Source Systems

Operational & transactional data

02

Incremental Load

Scheduled, idempotent ingestion

03

Governed Warehouse

Star schema · governed views

04

Executive Portal

Boardroom-grade screens

05

AI Answer Layer

Read-only governed queries

Governance & security — applies to every layer above

Role-based accessRow-level securityAudit loggingRead-only query paths

Capabilities

Six layers that share one definition of the truth.

01

Executive BI

A dense, boardroom-grade portal rather than a scattering of one-off reports. Every screen answers a question an executive actually asks in a review.

  • Executive summary and network health overview
  • Trend and geography views with drill-through
  • Location leaderboard and terminal utilization
  • Game/product performance and math performance
  • Jackpots and progressives monitoring

02

Forecasting & Anomalies

Forecasts and anomaly detection sit on the same governed numbers as the dashboards, so a projection and a report never disagree.

  • Holt-Winters time-series forecasts with confidence bands
  • Network, market and location level horizons
  • Severity-ranked anomaly surfacing for triage
  • Deviation framed against expected range, not raw noise

03

Player Analytics

Behaviour analysis built as a reusable layer instead of an ad-hoc extract, so cohort and value questions are answerable on demand.

  • Cohort retention curves
  • Segmentation and churn-risk views
  • Player-value analysis across segments

04

Experimental Design

Promotion and configuration decisions are treated as experiments with a designed matrix, not as opinion contests.

  • Taguchi L18 style orthogonal design
  • Main-effects decomposition per factor
  • Signal-to-noise ranking of configurations
  • Promotion design informed by measured effects

05

AI Query Layer

Natural-language questions are translated into constrained, read-only queries and tools against the same governed layer — auditable answers, not free-form number generation.

  • Question → governed tool/query, never improvised SQL on raw tables
  • Answers reuse the same metric definitions as the portal
  • Every response traceable to a governed view
  • Tool-constrained surface limits what the model can reach
See it in Nucleus ↗

06

Governance & Security

Access and traceability designed in from the start, because an executive platform is only trusted if it is controlled.

  • Role-based access by function
  • Row-level security on scoped data
  • Audit logging of queries and access
  • Strictly read-only analytical paths

Engineering decisions

The four choices that make the rest of it hold up.

Why one governed metric layer?

Hold, RTP, rotation and net/store/day are defined once in SQL and governed views. Dashboards, forecasts and the AI layer all read the same definition, so KPI drift between a deck, a report and a chatbot answer becomes structurally impossible instead of a review-meeting argument.

Why incremental scheduled loads?

Full refreshes get slower and more fragile as history grows. Incremental, scheduled ingestion keeps load windows predictable, makes reruns safe, and keeps the executive portal current without a fragile overnight monolith.

Why a shared warehouse instead of per-dashboard extracts?

Per-dashboard extracts fork the truth. A star schema with governed views on top gives one modelling surface, so adding a screen or a statistical engine is a read against existing dimensions rather than a new pipeline.

Why tool-constrained, read-only AI?

An LLM given open database access is both a correctness risk and a security risk. Constraining it to governed, read-only tools means answers are reproducible, permissions still apply, and every number can be traced back to a definition.

What this project demonstrates

For hiring managers: the scope a BI or analytics leader is actually being asked to own.

Analytics architecture

Designed the path from source systems through incremental loads into a star schema with governed views — and kept the metric layer as the contract.

Statistical depth

Holt-Winters forecasting, anomaly severity scoring, cohort retention and Taguchi-style designed experiments implemented as engines, not one-off analyses.

Product thinking

Scoped screens around the decisions executives actually make, and sequenced the build so each layer made the next one cheaper.

Executive communication

Dense boardroom-grade views that lead with the answer, with drill-through available for anyone who wants the working.

AI orchestration

An answer layer built on governed tools and read-only query paths, so natural-language access is auditable rather than speculative.

End-to-end ownership

Data modelling, statistical engines, front-end UX, access control and AI integration — designed and built as one coherent platform.

Technology & methods
  • TypeScript / React
  • SQL
  • Warehouse & star-schema modelling
  • Governed metric views
  • Holt-Winters forecasting
  • Taguchi DOE
  • Cohort analysis
  • Anomaly detection
  • Role-based access & RLS
  • AI tool orchestration (MCP-style patterns)

For recruiters & hiring managers

Don't just read about it. Use it.

A fully interactive environment seeded with synthetic data and strictly read-only — safe to explore the governed metrics, forecasts, and AI query layer without touching anything real.

Happy to walk a hiring team through how Nucleus was designed — the modelling, the statistics, and the governance behind it.