Skip to main content

Declaration · 10 November 2026

The final business declaration

TITANOS is a one-person AI engineering practice in Brisbane, run with a team of AI agents and a strict test gate. It builds software, security checks and AI systems for small and mid-sized businesses.

This page is everything in the open: what has been built, what it cost, how fast it ran, how it is tested, how it is kept safe, and how you can do the same. It is free to read, copy and share.

Key achievements

AI engineering capability

Cost

  • About US$30 a month of AI tools while the main engine was built (to 29 September 2026).
  • About AU$300 (estimate, source pending) a month of AI tool subscriptions since then.
  • Cost per AI job fell 8.4 times, from US$0.97 to US$0.12, while quality on the matched set rose from 23 of 25 to 25 of 25.
  • Receipted subscription costs are not measured yet; the two monthly figures above are the operator's own numbers and are labelled that way in the appendix.

Speed

  • About 168 commits a day on average across those 43 days.
  • 1,549 agent jobs in one 24 hour window, with a 7.3% failure rate (failed or stuck) and a median job time of 12 minutes.
  • Case studies below give start and finish times for whole builds, taken from the git log.

Quality

  • Nothing counts as done until the full gate is green and the result has been read, not assumed.
  • A second agent reviews each change; no agent signs off its own work.
  • Cost per AI job fell 8.4 times, from US$0.97 to US$0.12, while quality on the matched set rose from 23 of 25 to 25 of 25.

Tests and automated testing

  • 527 automated test suites in the engine's main gate.
  • 149 test commands in this website's test manifest, plus built-page checks that run after every build.
  • 114 test files and 1,054 test cases in the customer backend (sign-in, billing, security).
  • 21 mutation-test scripts that break the code on purpose to prove the tests notice.
  • Every failure that reaches a person becomes a regression test the same day.
  • Copy is linted too: no dashes, no filler words, and no addresses or credentials in anything public.

Security

  • Outside parties are only ever checked passively, from public DNS records. Nothing of theirs is touched.
  • Money, credentials, legal commitments, messages to third parties and anything irreversible need a human approval tap first.
  • Secrets live outside the code and are never pasted into chat; a secret scan is part of the engine's test gate.
  • The website sets a content security policy and publishes a security contact file (security.txt).
  • The customer backend verifies signed payment events, rate limits sensitive routes per visitor, and sends no-store and anti-framing headers on every response.

Hardware

  • The workstation is a laptop: Intel Core i7-7700HQ (8 logical CPUs), 23 GB of RAM, running Linux under Windows.
  • Full test suites run on a cloud test runner: AMD EPYC 7B12, 4 cores, 15 GB of RAM.

Subscriptions and prices

Prices are read from the site's own price file and link to the page that sells each one. See also /pricing and /costs.

Case studies

Used vehicle dealership · REAL

A dealership command centre, built and independently reviewed in about 70 minutes

  • About 70 min from the first job starting (15:36) to the reviewed build (16:46), 14 October 2026
  • 23 jobs run in parallel: 4 discovery, 10 builders, 8 reviewers and 1 lead
  • 598 / 598 automated tests passing
  • 12 ms to list 10,000 vehicles (slowest 5% of runs, one database measurement)
Read the case study

Sales coaching · REAL

A 1,900-lesson sales course with its own coach, built in under 90 minutes

  • 83 minutes from the first commit (10:44) to the full course (12:07), 9 October 2026; 409 of its 412 commits
  • 1,900 lessons in 190 modules, about 2.7 million words, each lesson with a five-question quiz
  • 412 commits in the project history
  • 35 / 35 automated tests passing
Read the case study

Education and practice trading · REAL

A full trading-education app, idea to working product in one morning

  • 3.5 hours from first idea (05:10) to the finished app (08:40), 7 October 2026
  • 56 commits in that window, all on the record
  • 250 / 250 automated tests passing
  • About 6% of one week of a flat-rate AI subscription
Read the case study

Each case links to its own page with dates, costs and what was tested. Client names are never published, and every figure traces to the build's git log or a dated usage reading.

How we do it

  1. Write down what done means first. Before anything is built, list the acceptance checks: what must work, what must not break, what the limits are. Everything later is judged against that list, not against a feeling.
  2. Run one orchestrator and many small agents. One session holds the plan and the context. It hands self-contained briefs to agents, each with one job, its own branch and the exact files it may touch. Small, bounded jobs fail cheaply.
  3. Tests before claims. A change is not described as working until a test shows it. Each bug that gets through becomes a regression test, and mutation tests confirm the tests really fail when the code is wrong.
  4. Run the full suite on a cloud test runner. The workstation stays free for building. The whole suite runs on a separate machine from a clean copy, and the pass and fail counts are read, not assumed.
  5. Use independent review. A different agent, with no stake in the change, reviews it. It only blocks on a demonstrated fault. Everything else goes on a list for later, so work keeps moving.
  6. Put approval gates around the irreversible. Spending money, handling credentials, signing, messaging a third party and deleting things go to a person as a one-tap card. Everything reversible and in scope just gets done.
  7. Keep receipts. Every change leaves a commit, a test result and a note of what is still unknown. A number without a source and a date does not get published.
  8. Turn repeats into code. Anything done twice becomes a script with a test. That is how a one-person practice keeps pace: the machine does the known, the person decides the unknown.

This is not magic and it is not hands-off. A person sets the direction, approves the gated steps and owns every decision that matters.

Courses

Trading platform (includes the AI course)

A phone-first trading education platform: a 200-lesson course, live charts, an AI tutor on every screen and a practice-money journal. The course on how to use AI sits inside it.

Sign up

See how it was built

Sales course

A sales course of 190 modules and 1,900 lessons with quizzes and an AI coach for questions, role-play and speaking practice, built for a sales coach.

Join the waitlist

Daily artifacts from TITANOS. Unsubscribe any time.

See how it was built

Daily artifacts

One short, useful artifact a day from the build: a number, a method, a template or a lesson, with the source attached.

Daily artifacts from TITANOS. Unsubscribe any time.

Appendix: sources for every number

Every figure in this declaration, with where it came from and when. Measured on 10 November 2026 unless a different date is shown. MEASURED means counted from a real source. COMPUTED means arithmetic on measured figures. PUBLISHED means read from the live price file. OPERATOR_STATED means the operator's own figure with no receipt on file. Figures dated 9 October 2026 or 13 October 2026 come from the site ledger and were not re-measured for this document.

FigureStatusSourceDate
7,236MEASUREDgit rev-list --count origin/main, engine repository. Commit count on the main branch; first commit date from git log.2026-11-10
168COMPUTEDengine_commits divided by 43 calendar days (29 September to 10 November inclusive). 7,236 / 43 = 168.3.2026-11-10
1,041MEASUREDgit rev-list --count origin/main, website repository. Commit count on the main branch; first commit date from git log.2026-11-10
527MEASUREDdistinct test_*.sh entries in the SUITES list of the engine's verify script on origin/main. Entries counted by script, duplicates removed (528 listed, 527 distinct).2026-11-10
149MEASUREDnon-comment lines of scripts/test_manifest.txt, website repository, origin/main. Line count of the manifest the test runner reads.2026-11-10
1,054MEASUREDportal-api/test/*.test.ts in the website repository. Files counted with ls; test() and it() declarations counted with grep (declarations, not executions).2026-11-10
21MEASUREDscripts/test_*mutants.sh, website repository. File count.2026-11-10
6,314MEASUREDsite ledger (lib/ledger.json, tests_main), from the engine's recorded verify runs. Recorded gate results. Not re-run for this document.2026-10-09
256,217MEASUREDsite ledger (lib/ledger.json, estate_sloc); method on titanos.tech/engineering. Non-blank, non-comment lines; mirrors, clones and third-party code excluded. Not re-counted for this document.2026-10-13
8.4xMEASUREDsite ledger (lib/ledger.json, cost_job_x), from the kernel compression gate result. 25 matched jobs run before and after the compressed instruction set.2026-10-09
1,549MEASUREDsite ledger (lib/ledger.json, fleet_jobs_24h, fleet_fail_rate_24h, fleet_median_job_min), engine job ledger. Job ledger totals for the 24 hours to the stated time. Not re-measured for this document.2026-10-13
8 cores, 23 GBMEASUREDnproc, free -g and lscpu on the workstation. Commands run on the day.2026-11-10
4 cores, 15 GBMEASUREDnproc, free -g and lscpu on the cloud test runner. Commands run on the day. The runner's address is deliberately not published.2026-11-10
about US$30 a monthOPERATOR_STATEDtitanos.tech/engineering, operator figure; no receipt on file. Stated by the operator. Not receipted.2026-10-09
about AU$300 (estimate, source pending) a monthOPERATOR_STATEDtitanos.tech/engineering and /costs, operator figure; no receipt on file. Stated by the operator. Not receipted. A receipted subscription ledger is not measured yet.2026-10-09
AU$0 to AU$149,997PUBLISHEDlib/pricing.ts in the website repository. Read from the file the live site renders its prices from.2026-11-10

Case study figures: taken from lib/caseStudies.ts as at the date above. Each is traceable to that build's git log or a dated usage reading, and the case page shows the detail.

Not measured yet: a receipted subscription ledger, lines of code re-counted after 13 October 2026, and any revenue figure. They are left out rather than guessed.

Book a free call