Nordwind Rollout — margin down 4.2 pts against baseline
- metric
- gross_margin_pct
- value
- 21.8%
- threshold
- 26.0%
- Δ
- −4.2 pts
Open project · costs
The financial close for project businesses
Signals runs the money side of delivery — cost sheet to purchase order to invoice to timesheet to a governed monthly close — and keeps watching between closes. Margin erosion, cost overrun, schedule slip and cash risk are detected by domain math, ranked, and handed to the person who can still do something about them.
Open project · costs
Open receivables · aging
Open project · earned value
01 — the gap
A portfolio review is a snapshot. Between two of them a burn rate doubles, a milestone quietly slides, an invoice ages into the 90-day bucket, a resource ends up committed at 140%. All of it is already in the system — in the closure, the timesheet, the supplier invoice, the allocation matrix.
Nobody is short of data. What is missing is something that reads it on every tick, decides what actually became material, and says so while it is still cheap to fix.
The month-end number is a verdict.
A signal is a warning.
02 — architecture
An intelligent system is three parts, and capability is dominated by the first two. Most products in this space are a model with a thin wrapper. Signals is the other way round: the hard, durable part — a full portfolio management system with the domain math already in it — is the part that was built first.
01
The running system. World state and domain expertise.
A live database of projects, cost sheets, earned value, closures, capacity, invoices and acceptance — plus the deterministic detectors that compute over them. PMO practice encoded as arithmetic rather than as prompts.
Durable · model-agnostic · already built
02
The iteration that turns state into intelligence over time.
Sense, detect, reconcile, explain, rank, recommend, record, observe the outcome, adapt. Continuous rather than requested. This is what makes the system notice things instead of answering questions.
The compounding part
03
A reasoner plugged into one step. Interchangeable.
The model writes the explanation and helps with priority. It never manufactures evidence. Swap it, downgrade it, or configure none at all — the deterministic signals still fire, with templated explanations in your language.
Commodity · swappable · optional
Which is also the risk answer: nothing here is a bet on one vendor's model, and none of your portfolio data has to leave your infrastructure to make it work.
03 — the loop
The sweep runs on a schedule against the same service layer the application uses. It is convergent: re-running it never duplicates a signal, and a condition that has cleared resolves itself on the next pass.
Signals recommend; people decide. The loop holds read scope only — it never edits a project, a closure or an invoice. Every recommendation is a link into the one surface that owns that change.
A candidate has to clear a configured materiality threshold before it is ever raised. The model may demote a signal; it cannot promote one into existence. Cleared conditions expire themselves.
Detection is threshold and earned-value arithmetic on your live data — not a model trained on someone else's portfolio. It is useful on day one and explainable on day one.
04 — evidence
The line between what is computed and what is written is drawn in the data model, not in the marketing. Evidence is deterministic and reproducible: same inputs, same numbers, no model involved. Only the prose varies with the reasoner.
Remove the model and rows 4–5 fall back to a templated explanation. Rows 1–3 and 6–7 do not change.
05 — coverage
Nine signal families, each wrapping a detector that already runs against your portfolio. Thresholds and severity tiers are configuration, per instance — not constants baked into a vendor's idea of a healthy project.
06 — the book of record
Detection needs a plan to compare against, actuals to compare with, and a governed month-end that decides which numbers are final. Tracking budget against actuals continuously is table stakes — producing a frozen, approved record of each month, reconciled, drift-disclosed, and hashed into a chain, is not. That close is the product, and it runs a delivery organisation on its own.
agent-native
Everything the interface can do is also reachable over a command-line client and a Model Context Protocol server, under a real user's token and the same permission model. Point your own assistant at the portfolio: it gets exactly the access that person has, and every write lands in the audit trail with a name on it.
07 — who it is for
Portfolio-wide exposure, a closure pipeline that chases itself, approvals with an audit trail — and a month-end record Finance can stand behind.
One project cockpit: plan versus actual, what is blocking the close, what needs deciding this week — scoped to the projects you own, and nothing else.
Timesheets, leave, and the work assigned to you. Small surface, few clicks, no access to anyone else's numbers.
Next
A walkthrough runs on a demo estate first, then on a copy of yours. Self-hosted on your infrastructure, or run by us — the data model and the detectors are identical either way.