Asset
The atomic learnable or deliverable unit — a procedure, a piece of knowledge, a technique, a tool-use. The smallest thing you can actually build or demonstrate.
Docs
Under the hood.
Capabilities, architecture, the ontology, and where the data comes from. SMART is a real, runnable system — this is what's actually inside it.
Capabilities
What the system does.
- Per-individual capability graphsLive proficiency, confidence, evidence and decay per person, measured against role requirements.
- Three AI copilotsCapability Mirror (assess), Personal Coach (RAG-grounded, SOP-cited guidance + microlearning), Career Navigator (prerequisite-ordered route to a target role).
- JD → training pipelineDissect a job description into required skills, assemble the ideal evidence-backed set, gap-analyse against an incumbent, emit a cited training plan.
- Ideal-workforce design & scenariosModel the capability distribution a function should have and test interventions against it.
- Knowledge & bus-factor analysisSurface concentration risk, capture knowledge into reusable assets, model the cost of a departure.
- Profit-lever mappingSkills linked to named profit levers so capability connects to business outcomes.
- PPI-for-People analyticsCoverage, engagement, intervention effectiveness, and real AI token/cost captured per copilot call.
- Multilingual UI43 languages with flagged English fallback for untranslated strings.
Architecture
TypeScript end to end, one deployable.
A React + Vite single-page app talks to an Express + SQLite server over a small JSON API. Model keys live server-side and never reach the browser.
# web → /api → server (single deployable)
web/ React + Vite + Recharts SPA
│ /api
▼
server/
├─ engine deterministic core — gaps, prereqs, decay, coverage
├─ copilots the 3 copilots — real LLM calls, structured IO
├─ rag SOP retrieval — vector embeddings, lexical fallback
├─ evidence-bank graded / cited / current evidence packs
├─ evidence-rag retrieval ranked by tier × match × confidence
├─ memory longitudinal capability + intervention effectiveness
├─ pipeline JD → dissect → ideal set → gap → training plan
├─ measurement event pipeline + analytics aggregation
└─ llm provider-agnostic, OpenAI-compatible client
Provider-agnostic by design. The LLM client speaks the OpenAI-compatible protocol, so the chat and embedding models are swappable via environment configuration. With no key set, the system stays fully runnable on deterministic, clearly-labelled offline fallbacks.
The ontology
Assets → skills → roles.
The ontology is the data spine. It's compositional, queryable, and additive — new skills can be minted from real work without breaking what's already modelled.
Skill
A composition of assets plus an evidence pack describing what builds that capability best. A skill is never just a label — it carries the proof of how to develop it.
Role
A composition of skills with proficiency targets. Roles are what JDs map onto and what gap analysis measures a person against.
The graph is queryable end to end. In the current demo dataset that's 27 assets, 28 skills and 8 roles — and it grows as the pipeline mints novel skills from real job descriptions.
Data provenance
Where the data comes from.
The evidence model
Every skill carries a graded evidence pack. Claims are tiered — peer-reviewed > practitioner > grey — and each carries a source, a date and a confidence. Packs carry a last-reviewed date, and evidence-RANK ensures strong peer-reviewed evidence outranks weak grey at equal textual match.
Synthetic by design
The demo runs on synthetic personas and data. No real Thermo Fisher Scientific staff data is used — the system is modelled against representative operations roles to show the mechanism without touching anyone's records.
Organisation model
The demo is populated against a real, public organisation model — segments, hierarchy, corporate functions and global headcount across 50+ countries. Country-level headcount is a clearly-flagged medium-confidence estimate derived from facility sizes, revenue-share proxy and public location data, since the figures aren't published.
Lever economics
Profit-lever figures in the demo are sourced from the modelled organisation's public 10-K filings. They illustrate how capability connects to financial levers — they are not claims about outcomes SMART itself has produced.