Series
Human in the Loop
Field notes on managing AI-assisted software delivery: how I scope the work, what I check before anything ships, and the bugs that got through anyway.
Published
5
Planned
5
Every post is grounded in a real repository: commit hashes, measured numbers, and the gaps where I didn’t measure.
The series in one picture
Most of the work is deciding what gets through
live simulation
Candidates
0
Accepted
0
$ / accepted
—
Illustrative: $0.06 per attempt. The cheap part is the attempt; the real unit is what gets accepted.
Published parts
- Part 10 Read →
Sep 14, 2026 · unalloc
unalloc: checking a cost-attribution tool before anyone quotes its numbers
Nine defects caught before release, a $4.41-an-hour GPU with a watchdog, a paper build that refuses silent typos, and a headline softened from "over-charging" to "the rules disagree".
- Part 09 Read →
Sep 14, 2026 · timurista.ai
timurista.ai: rebuilding this website in a day, with sign-off at every step
A framing pulled within the hour, numbers deleted for lack of a source, a form that showed my callback number to the wrong people, and the checks that caught them.
- Part 03 Read →
Sep 13, 2026 · TigerMill
TigerMill: quality gates for an AI content pipeline
Acceptance rules written before looking, attributed approvals, and why the metric that matters is dollars per accepted panel.
- Part 02 Read →
Sep 13, 2026 · TrendVesting
TrendVesting: validating 2.5 years of production AI
Invariant tests, reconciliation audits, and calibration checks that overturned numbers I had been quoting.
- Part 01 Read →
Sep 13, 2026 · TendForm
TendForm: managing a solo SaaS build to production quality
Scoping, type-checked deploys, verified-live checks, and the edge cases that still slipped through a HIPAA tier.
Planned
Working titles. Part numbers are reserved, so new parts can land out of order.
- Part 04Planned
Governance
Your permission allowlist is an org chart
How allowlists, rules files, and deploy runbooks turn one-off approvals into policy, and where they quietly over-grant.
- Part 05Planned
TendForm · TigerMill
Instructions rot faster than code
Project docs that still describe infrastructure replaced on day one, and a cheap review habit that keeps them true.
- Part 06Planned
TendForm · TrendVesting
MCP is a product surface, not an integration
Designing tools that are hard to misuse: rules inside tool descriptions, preview links instead of images, and no spend in the toolbelt.
- Part 07Planned
Unit economics
Dollars per accepted output
Measuring AI-assisted work like a cost center: rework, rerolls, review time, and the rejection rate nobody tracks.
- Part 08Planned
Workforce CLI
Running parallel work streams without losing track
One lead per repository, task notes that survive restarts, and when a single focused session beats five.