Finance decision tools, built with AI — for founders, finance teams, executives, and boards, from early-stage startups to global technology companies. Working examples of one repeatable motion: identify the operating question, define the finance logic, build the tool, translate the result into a decision. These are portfolio demonstrations, not one recommended suite for a single company — each card states the stage it fits, who runs it, and the decision it answers. Every project ships a validation step that must pass before its dashboard publishes.
Jump to: Executive & board · Founder operating tools · Business-model-specific · Accounting, controls & close
What boards and investors actually read — the statement model, the scenario distribution, and the metric pack, with every definition on the page.
Thirty metrics in five sections — growth, customer health, efficiency, financial, AI-specific — every one with its definition on the page. LTV:CAC is flagged as formulaic right on the scorecard.
NRR, GRR, burn multiple, CAC payback, magic number, Rule of 40 — each computed by a function whose docstring states which definition it uses. The ARR walk has to tie, and the report says whether it did.
Income statement, balance sheet, and indirect cash flow, line by line across twelve quarters with scenario tabs. A = L + E to the cent in every column, or the model refuses to publish.
10,000 simulations of 12 quarters. P10/P50/P90 fans for ARR and cash, floor-breach probability, and a tornado ranking which assumption actually drives the outcome.
The internal operating cadence — where revenue is heading, what pricing really realizes, what the team costs, and how long the cash lasts.
Next four quarters from the three books that exist today: sales capacity, stage-weighted pipeline with coverage flags, and a per-contract risk-weighted renewal book.
Cohort heatmap, layer cake, and empirical CAC-payback curves by channel. Seeded with a promo cohort that dies at first renewal — invisible in blended NRR, unmissable in the heatmap.
List → pocket price across 1,100 deals. Quarter-end deals realize 5.9 points less — ~$810K given to the calendar — and discount depth creeps with deal size. The waterfall ties on every deal.
The hiring plan is the budget. Fully-loaded cost runs 34% over the salary-only version, and 28% of approved heads aren't productive capacity — req-to-start lag plus ramp, not a recruiting failure. Three hiring scenarios.
Gross vs net burn, runway trajectory, and the fundraise trigger drawn as a date on the chart. Runway is measured on operating burn — a trailing average can't absorb a one-time prepay without printing a cliff that isn't real.
Direct-method forecast from the AR aging, the AP aging, and the recurring commitments in neither. Low point, covenant floor test, and collections stress scenarios.
Different business models break generic finance tooling in different places. These are built for the economics that make each model unusual — hardware as a capital asset, inference as COGS, revenue in basis points.
A per-robot P&L across 181 deployed units: utilization vs uptime, contribution against the depreciation each unit must out-earn, and payback per unit. 18% of the fleet never repays its hardware — for two different reasons that need opposite fixes.
Is a training run R&D expense or an intangible asset? Both treatments modelled side by side, with the identity that keeps them honest: over a run's full life they charge the same total. I predicted the EBITDA gap would compress; validation proved it compounds.
True CAC counts the pilots you lost — $580K, 5.3× the naive figure. Conversion collapses past nine months, making pilot purgatory datable. Run only qualified pilots and CAC falls to 31% of three-year TCV.
Two questions kept apart: pricing (same 60 accounts and 195M tasks, four invoices) and distribution (OEM reach, where demand deliberately isn't constant). As a price royalty is worse; it wins only on reach — and that's where 61% concentration comes from, not from the rate.
Cost per successful task — with task-success and human-intervention guardrails beside every cost claim, because cheaper at degraded quality is a regression, not a saving. A model-routing deploy cuts cost per successful task 41% with quality held flat; tokens and GPU-hours sit underneath as drivers.
License + metered usage against committed minimums. A run-rate revenue walk (committed / variable / one-time split, with the metered event defined), a commit-utilization action list that excludes still-ramping deployments, and the rate-compression curve.
50,000 transactions matched against a bank settlement file. Every break classified — timing, amount, missing, duplicate, FX, fee — aged by exposure and routed to an owner. All 1,148 seeded breaks found and correctly labeled.
Contribution per transaction across 120,000 transactions by rail, region, and merchant segment. Contribution waterfall, cross-border premium, and mix-shift sensitivity.
The reliability layer — revenue recognized correctly, spend monitored continuously, consolidations that tie, commentary generated from computed figures. The part of finance that makes the charts trustworthy.
US parent + three foreign subs: translation vs remeasurement, with CTA computed two independent ways every month — a plug always balances, which is exactly why it can't be the only computation. The intercompany matrix catches a seeded ¥27,000 break and flags it as a transposition: the difference divides by 9.
A 26-lease register where every schedule must amortize to exactly zero and the footnote must tie to the liability. The largest lease never uses the word: a $10.1M "GPU capacity services agreement" with dedicated racks — the embedded-lease question every AI company's compute contracts are quietly asking.
A grant ledger where the expense engine must match an independent closed-form recomputation at 96 month-ends. Straight-line vs graded side by side, a RIF that reverses in-cliff expense at actuals, and a repricing costed the way 718 demands: incremental fair value only — an underwater option still cost what it cost.
Standalone-selling-price allocation on 40 multi-element bundles: "free implementation" gets revenue anyway — you can discount an invoice line, not the accounting. Plus the principal-vs-agent call, scored on the 606 indicators and proven by arithmetic: gross would show 6× the revenue with identical profit.
Six closes as a dependency graph: days-to-close falls 8.9 → 6.9 by automating the critical path's longest task — then the dashboard names the next bottleneck. A flux review with materiality thresholds refuses to certify November until one unexplained breach gets its commentary.
Bookings → billings → revenue from a 857-contract ledger. Deferred and RPO roll-forwards tie to the cent every month, with the deferred unwind schedule and RPO disclosure split.
Seven audit detectors over a year of AP and T&E — duplicates, split invoices, Benford's law, policy and velocity anomalies. 100% recall against seeded ground truth, false-positive rate reported.
Budget-vs-actuals in, board-ready variance commentary out. Python computes every figure; Claude drafts the narrative. The model never does arithmetic and never sees a number it wasn't handed.
--validate
step scoring the analysis against known ground truth. Where a check
disagreed with what I expected, the README says so. Source and full
write-ups at github.com/Lumimama/finance.