What a Team Would Have Billed (Part 4 of 6)
Three independent estimation methods converged on an uncomfortable number — roughly 40 person-years of engineering, delivered in 4.4 months.
Series: Solo + AI Delivery Efficiency — Part 4 of 6.
What a team would have billed
I ran three independent estimation methods on the codebase. They converged on an uncomfortable number: roughly 40 person-years of engineering effort — delivered in 4.4 months.
The surface area: ~303,000 lines across 38 domains. 2,400+ Elixir modules, 13,000+ public functions, 393 LiveViews, 324 data schemas, 4,000+ test cases.
Three ways to size that:
- COCOMO (industry estimation model): ~35–80 person-years
- Productivity benchmark (real-world lines/day, including tests, review, debug): ~37 person-years
- Domain-by-domain sanity check: ~30 person-years
Central estimate: ~40 person-years embedded. Conservative band: 30–45.
A note on rigour and its limits: COCOMO isn't a back-of-envelope guess — Barry Boehm built it in 1981 from regression analysis on 63 real projects, and it's been an estimation standard since. It's also famously imprecise; accuracy studies put its average relative error near 100%. That's exactly why I quote a range, not a single confident figure.
And one honest framing I keep returning to: person-years are effort a team would bill, not headcount you must hire. Nobody needs 40 people. But even the floor of that range, delivered solo in under five months, is worth sitting with.
Sources
Figures are paraphrased; verify against the originals before quoting.
- COCOMO / Barry Boehm (1981), organic-mode formula and 63-project basis: datacamp.com
- COCOMO accuracy evaluation (~100% mean relative error on the NASA dataset): journal.fkpt.org
Previous: Part 3 — The Cost Inversion · Next: Part 5 — Before You Quote the 8×