Case studies

Four engineering stories from Falkor

Each follows the same arc: problem, evidence, investigation, decision, fix, verification, lesson. Every number comes from Falkor's own dated records.

Execution-grounded claim verification

Make the AI admit what it didn't do

Published · 5 min read

Falkor's chat tools were silently switched off on every real turn, because the system prompt contained the word “Falkor”. The fix went further: each answer's claims are now checked against what the turn actually executed.

My role
Owned the architecture and the acceptance criteria for the repair
Key decision
Switch behaviour on structure, never on text, and bind every claim to the turn's execution record
Result
A fabricated claim caught and corrected live; a real reminder created and verified through chat
5,068 → 502prompt tokens when the tools silently vanished
  • AI governance
  • Tool use
  • Approvals

Preventing false-positive certification

Green is a claim, not a fact

Published · 5 min read

A certification harness that could declare success without evidence. A watchdog accusing a healthy store. 136 information sources shown as “active” that nothing ever ran. This is how Falkor learned to prove its own status.

My role
Set the rule and owned the certification gates
Key decision
Every status comes from evidence about its subject, and every gate must fail before it is trusted
Result
19 negative controls, and on 16 Nov 2026 a full certification run passed all 8,724 browser tests
6ways certification could pass without evidence, all closed
  • Quality engineering
  • Observability
  • Release governance

Governing AI-assisted software delivery

One builder, a fleet of AI agents

Published · 5 min read

More than 5,700 commits since March, most of them written by AI coding agents working in scoped lanes, with handoffs, evidence gates, independent review, and an operator who overturns a “PASS” that isn't one.

My role
Designed the lane system and made the release decisions
Key decision
Treat agent output as a claim to verify, not as the product
Result
More than 5,700 commits since March, and a 649-capability audit checked before it was trusted
7 in 10main-repository commits signed by an AI coding agent
  • AI-assisted engineering
  • Program ownership
  • Agent orchestration

Stopping a rewrite and rebuilding in place

The rewrite that rebuilt Falkor

Published · 2 min read

A from-scratch replacement demoed with 0 of 13 user journeys working. It was frozen, audited for its best parts, and Falkor was rebuilt in layers instead.

My role
Commissioned and judged the rival designs, then directed the rebuild
Key decision
Freeze a rewrite whose demo ran 0 of 13 user journeys, and keep its best parts
Result
Falkor, checked the day the rewrite was frozen, was all green; the rewrite's best parts were kept
184 → 98page routes in two days, with no capability lost
  • Program ownership
  • Architecture
  • Delivery judgment

Working on something similar? Discuss a project, or see how I work.

Esc