Workforce Intelligence Platform
A workforce intelligence and transformation platform for a global bank's delivery organisation — linking client, account, project and employee so leaders can see talent gaps, utilisation, commercial exceptions and automation opportunities, then track what was actually done about each.
Connecting workforce data to the delivery hierarchy so leadership can see capability gaps, act on them, and prove where every number came from.
The platform models the real delivery chain — Client → Account → Project → Project Code → Employee — so every workforce metric can be traced to the commercial commitment it affects.
Its actual value is not any single screen. It is that work moves between people without falling through a gap: a change is proposed, an owner accepts or declines with a reason, and the record reflects reality only after approval.
Provenance is enforced structurally. Every figure carries its source, and a row cannot claim to come from a system that has never delivered data — that is prevented by test, not by convention.
Recommendations are deterministic and explainable by deliberate architectural decision. They propose; they never write business state and never auto-decide.
Workforce information sat in systems that never spoke to the delivery hierarchy. Leadership could see headcount, and separately see projects, but could not answer the question that mattered: where is capability thin against what we have committed to deliver?
The deeper problem was accountability. Recommendations were produced, discussed, and then evaporated — no owner, no record, no way to tell six months later whether anything changed.
And in an organisation this size, a number on a leadership screen gets acted on. If its provenance is unclear, decisions get made on figures nobody can defend.
A workforce intelligence and transformation platform for a global bank's delivery organisation — linking client, account, project and employee so leaders can see talent gaps, utilisation, commercial exceptions and automation opportunities, then track what was actually done about each.
If you need leadership to act on data rather than argue about it, this is the pattern: one authoritative model, provenance on every figure, and workflow that assigns an owner instead of producing another report.
Backend
126 modules · 180 endpoints
Frontend
43 screens · 136 modules
Data model
32 tables · 14 migrations
Tests
~11,000 lines across 69 files
Performance
Every read under 200ms at 100× volume
The constraint that shaped it
A ranking of people cannot be provisional
The talent-scoring mechanism is fully built but deliberately refuses to produce a result until leadership sets the weights. A provisional ranking of human beings gets read as a real one the moment it appears on a screen — so the system declines to render one rather than risk it. Refusing to answer was treated as a feature to protect, not a gap to close.
No generative AI — as a decision, not a gap
Recommendations use deterministic, explainable scoring. Calling arithmetic 'AI' would be exactly the overclaiming this product exists to prevent, and in workforce decisions an unexplainable recommendation is unusable. Generative capability is fenced behind an explicit architectural boundary and a feature flag for later.
Fabricated figures declare themselves
Where a data contract has not yet delivered, the platform shows the gap rather than inferring a plausible number. An early bug let records name a source system that did not exist, which also suppressed the warning banner — the fix made that structurally impossible.
Performance proven at 100× the demo volume
Benchmarked at 5,000 employees and 600 projects with no query count growing with row count — because the failure mode of an internal platform is looking fine in a demo and collapsing at real scale.
- Framed the product with leadership stakeholders and defined the delivery-hierarchy model
- Designed the end-to-end experience across 43 screens, including the handover and approval flows
- Established the architectural boundary governing what AI may and may not do
- Drove the provenance model — how the system communicates confidence and refuses to guess
- Hands-on across full-stack delivery, test strategy and performance validation
How We Solved the Problem
Hierarchy spine
The delivery hierarchy is modelled as the backbone every other entity attaches to, so workforce, commercial and capability data share one authoritative structure instead of four reconciled views.
Metric layer with declared provenance
Each metric family records whether it is derived from a real source, configured to an agreed business value, or awaiting a data contract — and figures without a live source declare that on screen and in the API rather than presenting as fact.
Recommendation ruleset
Explainable scoring surfaces capability gaps, low-impact work, mentoring and automation candidates — carrying its evidence, and retaining that evidence even when a human overrules it.
Handover workflow
Four governed handovers move work between employee, lead, leadership and delivery. Each transition requires a named owner to accept or decline, so nothing is left implicitly assigned.
Maker-checker permissions
Permission changes are proposed and approved by different people, and never by someone whose own access it touches.
Product Designs
These are real product screens from a client engagement. Enter the access code to view them.
NDA Protected
Concept designs available
Leadership home
Every figure on the executive view traces back to current data in the viewer's scope.
NDA Protected
Concept designs available
Capability gaps
Demand against available capability by skill, so thin areas surface before they bite.
NDA Protected
Concept designs available
Recommendations
Explainable recommendations that carry their evidence and wait for a human decision.
NDA Protected
Concept designs available
Talent ranking
Built, but refuses to produce a ranking until leadership sets the weights.
NDA Protected
Concept designs available
Maker-checker permissions
Permission changes proposed and approved by different people — never self-approved.
NDA Protected
Concept designs available
Data provenance
Where a source system has never delivered, the platform says so instead of inferring.
NDA Protected
Concept designs available
Action center
Accepted recommendations become owned work with a named owner and a status.
Back
View All Case Studies
Next Case Study
AI Assessment & Hiring Intelligence Platform
A multi-tenant enterprise SaaS platform that unifies assessment authoring, AI conversational interviews, resume intelligence, browser-based proctoring, AI answer evaluation and hiring analytics — built AI-first, but engineered to keep working when every AI provider is unavailable.