upload a timesheet or a brief. hudson traces every hour, invoice and reused asset to the margin it produced, finds the leaks, prices what to recover, and shows the repeatable work worth automating. nothing you upload leaves your browser.
every hour, every invoice and every reused asset, traced from the brief to the margin it produced. move the levers to test the outcome, the statistics under it, and the recoverable rand.
estimated against actual effort, per deliverable. the delta column is the estimation error the audit measures. cost of hours uses each role's rate, defaulting to the tenant baseline where none is set.
the numbers behind the verdict. estimation bias, the sensitivity of margin to each lever, the break-even the brief has to clear, and the recoverable range under a conservative reading.
actual against estimated hours, per deliverable. bars past the dashed line breach the 1.25× scope-creep threshold.
the marginal value of one hour equals the rate it bills at. cut the dearest hours first.
the firm floor is signed scope, collectable now. the band above it is creep the client may or may not accept as billable.
each deliverable at the rate it billed, sized by hours worked. the green line is the tenant baseline.
the engine flags six gap types. these are the ones live on this brief, ordered by rand impact.
the standard — the analysis is not the deliverable, the decision is
repeated technical outputs are not fixed cost. templatise them into skills and tasks, build once and run many, and those hours leave the cost base on every future brief. the reuse credit above counts reused assets; this counts reused capability.
cumulative margin from the factory across future briefs. the build cost is sunk at brief zero, the per-brief dividend accrues, and the line crosses into profit at break-even.
the automatable share of each deliverable, auto-classified by type. edit any share to test the dividend.
how it maps — analysis removes the toil, delivery banks the value
each step of the reconciliation, its method, and the figure it returns on this brief. the right column maps the step to the stage of the Insight Value Chain it belongs to, and the toolkit artefact that teaches it.
the same audit, explained for everyone from a first-time reader to a global tech lead. open the level that fits.
you gave hudson a list of jobs on a project, how many hours each was meant to take, how many it actually took, and what you charge per hour. hudson added up what those hours cost, compared it to what you invoiced, and worked out your profit margin.
then it looked for money leaking away: jobs that ran far over their estimate, hours billed below your normal rate, and scope you did the work for but never invoiced. it added up what you could still recover, and it flagged the repetitive jobs a computer could do next time so you stop paying people to redo them.
the one-line version: here is your real margin, here is where it leaked, and here is the rand you can get back.
cost of hours C = Σ (actualᵢ × rateᵢ), defaulting a missing rate to the tenant baseline. margin M = R − C, margin% = M ÷ R, tested against the 35% floor.
scope creep flags any line where actual > 1.25 × estimate; the billable overage is (actual − 1.25·estimate) × rate. rate mismatch is actual × (baseline − rate) where rate is below baseline. under-invoicing is SOW − invoiced.
estimation bias is the mean of the per-line signed errors (actual − estimate) ÷ estimate, reported with MAPE and standard deviation. break-even invoiced to clear the floor is C ÷ (1 − 0.35). the recoverable band runs from the firm floor (under-invoicing) to a ceiling that assumes every creep hour is disputed. the agent-factory dividend is Σ (shareᵢ × actualᵢ × rateᵢ) per brief, against a one-off build of Σ (fixed + 1.5 × automatable hours) × rate; break-even is build ÷ dividend.
this demo runs the exact reconciliation model that the Hudson service runs server-side. in production the engine is a FastAPI service on Azure Container Apps, computing over PostgreSQL with the Apache AGE graph extension: deliverables, hours, revenue lines and reuse edges are graph vertices, and margin is a traversal keyed on tenant_id under row-level security, so one tenant can never read another's trace.
the brief parser is an Azure OpenAI path with a hard cost cap and a fallback model; the reconciliation itself is deterministic arithmetic, so the numbers are reproducible and auditable rather than model-generated. the API returns the same JSON shape this page exports, which is what a technical evaluator should read first.
the agent-factory layer is the compounding argument: repeatable technical outputs are lifted into skills and tasks once, verified, and re-run, so their hours leave the cost base on every future brief. the payback is a standard build-once-run-many curve with an explicit break-even, not a productivity claim.
Keystone runs decisions through four stages. hudson is the same sequence, executed on one brief: define the margin decision, control the anchors, compute the trace, deliver the ranked call.
define the decision before the data. does this brief clear the 35% floor, and which lever answers it. the scoping chain runs objective → action → question → asset: protect margin, decide invoice-or-cap-or-re-rate, ask where the margin leaks, pull the time entries, invoices, SOW and rate card.
neutralise the anchors before they bias the read. the estimate and the baseline rate are anchors, not truths. the 1.25× line falsifies the estimate; the baseline test falsifies the rate. every deliverable is included, none dropped, so no selection gap hides in the total.
compute on clean data with the assumptions logged. revenue less cost of hours, per deliverable, no exceptions. the one assumption, that a missing rate defaults to the tenant baseline, is stated on the trace rather than buried, so a reviewer sees exactly what the numbers assume.
lead with the decision, not the metric. gaps ranked by rand, each carrying the margin it returns. the recommendation stack is answer-first: the largest recoverable action sits on top, phrased as the call to make, so leadership acts on it without reading the working.
tick the formats you want and set the order they matter to you, then generate. the branded PDF opens your print dialog; downloads ask you to confirm the save; copies land on your clipboard. nothing is uploaded.
the design-partner pilot puts hudson to work on your real briefs, with you shaping v1 and earning referral margin.