CRTCai

How CRTC AI builds · a true story

The night shift.

CRTC AI doesn't build with one chatbot. It runs a crew of five AI specialists — a system we call STARSHIP — that plan, build, and check each other through the night. Here's one real shift. Every number on this page links to a record.

Meet the crew

Five agents, each a different AI model, each with one job. The score next to each is its real track record from that shift — and yes, one of them is supposed to look bad.

Architect & Researcher

Vega

Reads the system, checks real websites, and decides what's worth building next.

Shift record: 5 passed, 0 failed · reputation 100
Builder

Forge

Writes the code and its tests, and repairs its own failures up to twice before giving up.

Shift record: 3 built, 1 failed, 3 blocked · reputation 71
Independent Verifier

Argus

Never builds. Re-runs every delivered build in a sandbox, re-checks its evidence, and tries to break it. Fails things on purpose.

Shift record: failed 4 of 6 it reviewed · reputation 46
Economist

Ledger

Tracks what every shift costs, what's real revenue, and keeps projections clearly labeled as guesses.

Shift record: 1 passed, 0 failed · reputation 60
Archivist

Scribe

Inspects the system at clock-in and writes the handoff and next-shift plan at clock-out.

Shift record: 3 passed, 0 failed · reputation 80
Why Argus scores lowest. A verifier that passes everything is useless. Argus's low score means it's doing its job — catching work that isn't ready, so a person never ships a broken build. That's the whole point of the crew.

One real shift

Overnight, November 4, 2026. 8 hours. Budget 50 cents of cloud AI (it ran 3 cents over, on credit we already pay for).

Read the system and three real local websites, then picked work worth doing.

Built a roofing estimate calculator and a water-heater sizer — real tools, with their own tests.

Re-ran both builds in a clean sandbox, re-hashed the evidence. Both passed.

A third build looked fine — but Argus found its tests missed a planted bug (a <= that should have been <). Failed it. It never shipped.

Wrote the handoff and queued the next shift's work. 112 events, all recorded in a tamper-proof log.

STARSHIP agent board showing each AI agent, its role and the jobs it ran
The real agent board from that shift. More in Proof of work.

Then the power died.

The same system runs on one laptop in Mount Pleasant. That morning, it lost power.

When it came back, 32 of its 33 services restarted on their own. The one that didn't, the AI diagnosed and fixed — and wrote itself a fix so the same thing can't break it again. We know because the recovery is logged: git intact, all 18 databases checked clean, 242 tests passing.

The crew is fast and it's relentless. But a person still approves anything that costs money, goes live, or gets sent. The AI does the legwork. A human signs off. That's not a limitation — it's the product.

See what we buildProof of work