01
Conflicting definitions
Enrollment offices count heads; planning counts FTE; research counts eligible ranks. Without a catalog, every dashboard invents its own truth.
University intelligence · MetaRef family
One answer. Every office.
Universities run on fragmented systems. UMIS puts enrollment, teaching, research, grants, and satisfaction behind one governed semantic layer — so the president, deans, and stewards see the same number, with provenance, every time.
Scripted 12–15 minute executive journey · Role-scoped authorization · Board-ready PDF
Same question. Same value. API, portal, BI, assistant, and PDF reports all route through one semantic contract.
Not an SIS. Not a SQL workbench. Not an autonomous decision-maker — a governed analytical layer for leadership.
The gap
Leadership keeps getting three different numbers
SIS, HR, finance, grants, LMS, and research systems each define “active student,” “FTE,” and “award” differently. Spreadsheets paper over the cracks — until a board meeting.
02
Scope escapes
A department head who can prompt an open chat model should still be refused salary data outside their unit — politely, and with an audit trail.
03
Unreproducible board packs
PDFs that re-query live data overnight diverge from the numbers leadership just reviewed. UMIS reports only from immutable query snapshots.
What UMIS does
A governed KPI layer with a grounded assistant
Fifteen demo KPIs (and a wider draft catalog) are defined once, versioned, authorized, and tested. Every interface — including voice — is just a client of that contract.
01
Semantic query contract
Frozen OpenAPI for metric_id, version, filters, comparisons, and suppression metadata. Hallucinated labels fail fast — with authz-scoped hints only.
02
President dashboard
Headline cards, five-year trends, college scorecard, and the Canvas evaluation finding — the hero screen of the live demo.
03
Text-first assistant + voice
Allowlisted tools only. Numeric claims must appear in tool results. Push-to-talk voice rides the same pipeline after grounding tests pass.
04
Suppression & audit
Primary and complementary small-cell masking fire on camera in the department-head beat. Policy refusals are logged for the steward view.
How it works
From source runs to a board PDF in four steps
1
Ingest & validate
Synthetic NIU CSVs load idempotently with checksums. Scorecard files are test oracles — never runtime tables.
2
Warehouse & marts
Dimensional core feeds semantic-ready marts. Golden tests assert scorecards match reference values to 2 decimals.
3
Authorize & query
Role + org scope + metric classification decide every row. The LLM never receives SQL or raw student records.
4
Report from snapshots
Conversation PDFs render only persisted query hashes — so on-screen and print figures reconcile exactly.
Who it's for
Built for the people who own the institutional story
Presidents & rectors
Institution-wide trends, comparable periods, and a board pack that matches what you just asked in the room.
Deans & department heads
Unit vs college vs university comparisons — without escaping your authorized scope.
Research & grants offices
Awards, success rates, and publications per research FTE from certified definitions.
Data stewards
Ingestion runs, quality results, KPI ownership, and an audit trail of every policy refusal.
MetaRef family
Scholarly quality tools that share one design language
UMIS sits alongside the publishing-integrity products on the MetaRef / CheckRef platform.
Questions
What executives ask first
Is this production university data?
No. This deployment is a demonstration using the synthetic Northgate International University (NIU) tenant. A persistent banner labels it as such on every page.
Can the assistant invent numbers?
No. Every displayed numeric claim must map to a returned semantic-query field or a deterministic calculation in trusted application code. Ungrounded claims are rejected by the verifier.
How is authorization enforced?
In the API, on every request, under the authenticated user's role and organizational scope — never by prompt instructions. Tool arguments that assert a different identity are ignored.
What about small-cell privacy?
The demo applies primary suppression (minimum cell size 5) plus single-query complementary suppression with deterministic tie masking. Full cross-query differential privacy is a productization item and is disclosed as such.
Is UMIS an SIS or ERP?
Explicitly not. UMIS does not perform admissions, employment, grading, funding, or disciplinary decisions. It is an analytical and reporting layer over governed definitions.
How do I present the demo?
Follow the scripted 12–15 minute journey: President dashboard → grounded text question → voice comparison → teaching drill-down → department-head suppression → authorization refusal → ambiguity handling → board PDF.
See the same number on screen and in the PDF
Open the NIU demo as President, switch roles on camera, and generate a board report from the queries you just ran.