BootUI is available at https://github.com/jdubois/boot-ui
An estimate of the effort that went into building BootUI with AI assistance, compared with an estimate of building the same project by a single developer without AI. Figures are derived from git history, GitHub pull-request metadata, and code metrics — not from time-tracking logs — and are intended as informed estimates rather than precise measurements.
Updated for the 1.0.0 release (2026-06-05). The final day added a VuePress documentation site, the redesigned Overview security & health scanner dashboard, Copilot token charts, a proxied-Hikari fix, a README/docs refactor, and the 1.0.0 release cut.
| Dimension | Measurement (v1.0.0) |
|---|---|
| Repository span | First commit 2026-05-25, 1.0.0 release 2026-06-05 → ~10.8 calendar days |
| Commits | ~264 total (on main) |
| Pull requests | ~223 (squash-merged, up to #239) |
| Authors | Julien Dubois (human driver) · Copilot agent (~44 commits) · dependabot · github-actions · 1 collaborator |
| Java | ~461 files / ~50,000 lines; ~81 test classes |
| Frontend | Vue 3 SPA: ~52 .vue components, ~40 feature panels, ~7k lines JS, ~35 Playwright e2e specs |
| Total tracked source | ~83,000 lines (Java, Vue, JS, XML, YAML, properties, Markdown) |
| Docs | ~5,800 lines of Markdown + a redesigned VuePress site published to GitHub Pages |
| Architecture | 5-module Maven build (core, autoconfigure, starter, ui, sample-app), Spring Boot 4 / Java 17, Maven Central publishing, full CI (build, CodeQL, release) |
It is a production-grade, multi-module Spring Boot starter that deeply integrates Actuator, Spring Security filter chains, Flyway/Liquibase, Hibernate, Micrometer/OTLP tracing, GraalVM reachability metadata, OSV vulnerability scanning, ArchUnit, and an embedded Vue SPA — roughly 40 distinct feature panels, each with backend endpoints, frontend views, and tests.
The project was built — through a tagged 1.0.0 release — in ~11 calendar days (25 May → 5 June 2026) by one human developer driving the GitHub Copilot coding agent. The cadence — ~223 PRs in ~11 days (~20 PRs/day) with an AI agent co-authoring commits — is impossible to achieve by hand. The estimate of actual human hands-on effort is ~80–110 hours (roughly 2 intense solo weeks), with the AI agent performing the overwhelming majority of the actual typing, scaffolding, and test writing.
- Velocity: ~223 merged PRs / ~10.8 days (~264 commits). The commit clock runs from ~05:00 to ~midnight across most days, consistent with a single person dispatching many asynchronous agent tasks in parallel rather than typing continuously.
- Authorship pattern: "Copilot" is a named commit/PR author (~44 commits)
alongside the human. The repo even contains
.github/copilot-instructions.mdand per-panel conventions — the workflow was explicitly agent-oriented. - The final day (1.0.0): even on release day the agent shipped a redesigned VuePress docs site, the Overview scanner dashboard, Copilot token charts, a proxied-Hikari detection fix, and a README/docs refactor before cutting the release — a volume of polish that would itself be several days of solo work.
- Where the human time actually went: writing task prompts, reviewing/merging ~223 PRs, resolving CI failures (Spring Boot 4 migration quirks, Flyway 11 API changes, OTLP property renames), and steering architecture. This is review-and-orchestrate time, not write-every-line time.
- Realistic human-effort band: ~80–110 hours. The ~11-day calendar window understates raw output (because the agent works in parallel) but the human reviewer/driver is the true bottleneck.
A single, experienced Spring Boot + Vue developer building the same scope by hand — no AI codegen, through a polished 1.0.0 release with a published docs site — would need an estimated 6.5 to 8.5 months of full-time work (~28–36 weeks, ~1,100–1,450 hours). The dominant cost is the ~40 feature panels, each requiring backend integration with a non-trivial Spring/JVM subsystem, a frontend view, and tests. Cross-check: a COCOMO "organic" estimate on ~50 KLOC of hand-written source yields well over 100 person-months — unrealistically high because much of the code is repetitive panel scaffolding — so a domain-expert solo figure of ~7.5 months is the defensible middle ground.
| Phase | Work | Est. time |
|---|---|---|
| 1. Project setup & architecture | 5-module Maven layout, auto-configuration skeleton, starter wiring, panel-registration framework, safety/access filter design | 1–1.5 weeks |
| 2. Frontend foundation | Vue 3 + Vite SPA, app shell, routing, grouped menu, shared components, build packaged into starter (no Node needed by consumers) | 1.5–2 weeks |
| 3. Core "easy" panels (~15) | Health, Metrics, Memory, Beans, Conditions, Mappings, Loggers, Config, Scheduled, Caches, etc. — mostly Actuator-backed. ~1.5–2 days each (backend + view + tests) | 5–6 weeks |
| 4. Complex / deep-integration panels (~20) | Spring Security filter-chain introspection, Security Advisor (37 rules), Pentesting (OWASP), Vulnerabilities (OSV), Hibernate Advisor, Flyway/Liquibase actions, GraalVM reachability generator, Threads/HeapDump, Tracing/OTLP, AI Usage, GitHub/Copilot dashboards, Architecture (ArchUnit). ~3–5 days each | 12–16 weeks |
| 5. Safety & security model | Local-only enforcement, action confirmation gating, secret masking, access filter — done carefully by hand | 1–1.5 weeks |
| 6. Testing | ~81 Java test classes + ~35 Playwright e2e specs, written manually alongside features | 3–4 weeks (partly overlapped above) |
| 7. CI/CD & release engineering | Build/CodeQL/release workflows, GPG signing, Maven Central publishing, javadoc/source jar plumbing | 1–1.5 weeks |
| 8. Documentation | ~5,800 lines: FEATURES, per-check catalogs, SPECIFICATION, properties, screenshots, plus a published VuePress GitHub Pages site | 2.5–3.5 weeks |
| 9. Integration, polish, Spring Boot 4 migration debugging, 1.0.0 release hardening, buffer | Cross-cutting bugs, version upgrades, regressions, release engineering | 2–3 weeks |
| Total | ~28–36 weeks ≈ 6.5–8.5 months |
| Aspect | With AI (actual) | Without AI (estimated) |
|---|---|---|
| Calendar time | ~11 days (to 1.0.0) | ~6.5–8.5 months |
| Human effort (hours) | ~80–110 h | ~1,100–1,450 h |
| Throughput | ~20 PRs/day, parallel agent tasks | A few features/week, serial |
| Human's role | Architect + prompt-writer + reviewer | Architect + author of every line |
| Bottleneck | PR review & steering | Typing, debugging, learning each subsystem |
| Test coverage | Generated alongside features (~81 + ~35 suites) | Same scope but hand-written, slower |
| Boilerplate (40 panels) | Near-free to replicate | Repetitive, the single biggest cost |
| Risk profile | Review fatigue; subtle AI-introduced bugs; needs strong CI (which exists here) | Fewer "surprise" bugs but far higher fatigue/abandonment risk |
| Effective speed-up | — | ~17–23× calendar, ~12–17× human-hours |
Why the multiplier is so large here: this codebase is unusually well-suited to AI assistance. It has (1) massive repetition — ~40 structurally similar panels — that an agent clones cheaply; (2) broad-but-shallow API integration across many Spring subsystems, where the cost without AI is mostly looking things up; and (3) strong guardrails (multi-module CI, CodeQL, e2e tests, explicit copilot-instructions) that let the human safely accept high agent throughput. Net effect: the AI compresses an estimated 6.5–8.5 month solo effort into ~11 calendar days and ~2 weeks of human attention — roughly a 12–17× reduction in human hours and a ~17–23× reduction in calendar time.
Most of the work was done:
- Using GitHub Copilot App on a MacBook Pro: this is for running a large number of agents in parallel, and testing the changes live before validating.
- Using the GitHub mobile app: this runs agents in Dockerized containers, allowing to test ideas and build prototypes on the go. This is where the AI workload estimate is a bit wrong: it adds many commits along the day, but that's because the developer was typing a quick idea on his phone. So the final working hours are probably a lot less per day than calculated here, and the working habits are clearly unusal for a developer.
The models used were:
- Mostly GPT 5.5 Extra High Reasoning, for about 1.5 milion tokens (90% cached)
- Claude Opus 4.8 and Gemini 3.1 Pro High Reasoning to complement it: for the trickest parts of the project, the developer used the 3 models to complement each other, and find the best solutions.
- Smaller model, or the "Auto" mode, for simpler tasks
Overall, token consumption was 120m input tokens, 15m output tokens, 2,500m cached tokens, for a total of 2,700m tokens and a total cost of about $2,000.
BootUI is a substantial, production-quality Spring Boot 4 starter (~83k lines, 5 modules, ~40 deeply-integrated panels, ~116 test suites, full release pipeline, published 1.0.0 with a VuePress docs site). With AI it was built — through a tagged 1.0.0 release — in ~11 calendar days and an estimated ~80–110 hours of human effort. The same scope without AI would realistically take a single experienced developer ~6.5–8.5 months (~1,100–1,450 hours).
The AI didn't merely type faster — it changed the developer's role from author to architect-and-reviewer, enabling ~21 merged PRs/day by running many tasks in parallel. The speed-up (~17–23× calendar, ~12–17× human-hours) is at the high end of what's achievable precisely because the project combines high structural repetition with broad shallow integrations and strong automated guardrails. The lasting lesson is that AI's biggest leverage is not on hard algorithmic problems but on large-surface-area, pattern-heavy, well-tested codebases — exactly this one. The human's judgment (architecture, the safety/security model, deciding what to build and what to merge) remained the irreplaceable bottleneck and the reason the result is coherent rather than just voluminous.
Caveats: estimates are derived from git/PR metadata and code metrics, not time-tracking logs; "without AI" figures assume one senior full-stack Spring/Vue developer already fluent in these subsystems and exclude calendar overhead (meetings, context-switching) that would lengthen real-world delivery.
Well done! One important detail: you know what you are doing and that makes the productivity grow very fast. Learn how to drive AI is the key skill at the moment! 👏 👏 👏