Skip to content

Case studies

Costs, at every scale

Enterprise cost attribution, and my own products run on invoices I pay myself. Same discipline at every size: find the unit that matters, measure it honestly, then change the system.

My own practice · 2025–2026

4 checks

between a proposed change and production

How I manage AI-assisted delivery without losing quality

The operating model behind three products I build and run solo: scoped work, gates with attribution, no spend in the toolbelt, verification before anything ships, and a written record of every decision.

Claude CodeMCPGit worktrees
Read case →

TendForm · Jun–Sep 2026

11 wks

first commit to a HIPAA-capable SaaS

A HIPAA-capable SaaS, shipped solo, priced against its own cost stack

A form builder with multi-form packets, a markdown DSL, a 27-tool MCP server, and a HIPAA tier, built and run solo in 11 weeks. The cost decisions and the verification habits mattered as much as the code.

React/ViteFastAPIKubernetes (DigitalOcean)
Read case →

TrendVesting · Feb 2024–present

+54.7 pts

stated confidence above the realized win rate

When cheap measurement overturned 2.5 years of beliefs about an AI product

A production AI signals platform I've run since February 2024. When I required every change to show its measurements, the numbers said the confidence scores meant nothing. That's the most valuable finding of the whole project.

GoTypeScriptPython
Read case →

TigerMill · Sep 2026

≈ $1

for the first 8 measured image-model calls

An AI content factory in 33 hours, and the metric that actually matters

A webtoon production pipeline with five gated stages, built over a weekend with acceptance rules written before any output was judged. Image generation is cheap per call; what it costs per accepted panel is the number nobody tracks.

GoSQLite queueReact / Three.js
Read case →

Roku · 2020–2021

$10–12M

AWS spend attributed per quarter

Attributing $10–12M a quarter of AWS spend, mid-migration

A cost-attribution pipeline for AI/ads forecasting workloads, built while co-leading the move from ECS/EC2 to Kubernetes. Cost data treated as an engineering artifact, not an accounting afterthought.

AWSKubernetesIstio
Read case →

Big tech · Big tech

~700K

servers under financial attribution

Making misattributed assets visible across a ~700K-server fleet

Financial attribution for private cloud infrastructure that runs internal AI workloads, plus internal platforms that make livability, deployment state, and cost visible to the people who own them.

Private cloudCost attributionInternal platforms
Read case →

Open source · Sep 2026

12–14 pts

disagreement between two metering rules on one tenant's H100 bill

An open-source AI spend ledger, checked on a real H100 before anyone quoted it

unalloc joins OpenCost allocations with LiteLLM, OpenAI and Anthropic bills and reports the AI spend nobody owns. Before trusting its numbers, I pushed six case studies through it, validated the metering results on a rented H100, and gated the paper's own build.

PythonOpenCostLiteLLM
Read case →

Method

How the work goes

  1. 01

    Attribute

    Map spend to owners, systems, and units. Nobody owns a cost they can’t see.

  2. 02

    Instrument

    Measure cost inside the system, next to quality and latency, instead of reconciling it after the invoice.

  3. 03

    Verify

    Scope the change, review it, and check it live: tests and invariants for the logic, a named human on anything subjective or paid.

  4. 04

    Make it legible

    Ship a surface finance, engineering, and executives can all read, and act on.

Want a number you can defend to finance?

Rates & fit check

Cash-pay advisory
from $250/hr

Check fit