01 · internal
Financial Engine
A bottom-up cash-flow model that reconciles to 99% of billed revenue.
99%
revenue reconciliation
0.90
AUC, payment prediction (1.0 = perfect)
73
golden parity tests
16.7k
lines of Python
/ interactive · drag the assumptions, the engine recomputes
try it, everything responds
NPV (6yr)
R7.8m
IRR (annual)
35.9%
breakeven
month 40
peak funding
R14.3m
cumulative free cash flow · 72 months
How to read this:you are building a number of network sites. Each one costs money to build now and earns a growing amount every month once it is switched on. The line shows cash in the bank over six years, starting deep in the red while you build and climbing as sites mature. NPV (net present value) is what all that future cash is worth in today's money. IRR (internal rate of return) is the effective annual return. Peak funding is the deepest the hole gets, which is the cash you must have to survive to breakeven. Drag any slider and every number recomputes.
A note on this demo. Every figure, name and data point here is invented for illustration. No client data, real economics or production configuration appears anywhere on this site. What is real is the machinery: the calculations, the rules and the logic running in your browser are the same ones the production system uses, so the demo behaves exactly as the real thing behaves.
What it is
/ in plain terms
The company builds WiFi networks in South African townships. This system predicts, month by month, how much cash each network will make or lose, then checks its own predictions against what customers actually paid and matches to within 1%. It answers questions like: if we spend R100 million building twelve more networks, when do we get the money back?
The financial brain of a South African township internet provider. A 16,700-line Python pipeline models every site, subscriber and voucher into monthly free cash flow, then derives the classic investment measures from it: discounted cash flow (DCF), internal rate of return (IRR), multiple on invested capital (MOIC), debt service coverage (DSCR) and breakeven. An actuals overlay reconciles the model against real billing to 99%, and a R100m capital-deployment simulator explores build-out scenarios. The engine was later ported to TypeScript with cent-level parity, proven by 73 golden-fixture tests.
Sole engineer. Data platform, financial engine, machine learning, dashboards.
The load-bearing details
- 01Monthly free-cash-flow model covering subscription and voucher revenue, site construction costs, bandwidth and rent, with conservative and refinancing valuation variants and the internal rate of return solved numerically by Newton's method.
- 02Actuals overlay reconciling the model to billing at 99% of billed revenue, on a single canonical definition of what counts as a subscriber.
- 03Self-supervised time-series transformer (a neural network that learns patterns from raw payment history without hand-labelled answers) trained on 95k devices across four years: 0.90 AUC against a 0.84 baseline. AUC, area under the curve, measures how well a model separates payers from non-payers; 1.0 is perfect and 0.5 is guessing.
- 04An out-of-time backtest surfaced a 12% conservative bias in the model's own calibration. It was documented rather than quietly corrected.
- 05Ported to TypeScript so it can run without Python servers. One implementation per metric, parity-tested to the cent.