Adopt a durable job queue
Background work was lost on deploys. The team compared a managed queue, database polling, and in-process retries before choosing a durable path.
- Owner
- Kai
- Why
- Reliable retries
AI FinOps · what your AI spend is buying
AI FinOps reads a provider export in this browser tab and names the one thing to act on: where to act first, how much is recoverable, what to do next, and how far the number can be trusted. The example decision below is already built and needs nothing of yours.
Your files do not leave this tab. Exports are read and analyzed in this browser. No upload, no credentials, no account, and nothing of yours is stored.
Built in this tab · nothing fetched, nothing uploaded
One question, one figure, one action, and the verdict that bounds them. Open the two levels inside it to check the benchmark, the confidence, the provenance, the limits, and the arithmetic — the same document, one keystroke deeper, rather than a second copy of it elsewhere on this page.
Printing or saving this page as PDF produces the briefing alone: every level opens on paper, and the site chrome, the controls, and the rest of this page come off.
The three synthetic periods behind it ship with this page, so there is nothing to fetch and nothing to wait for…
Your own numbers: analyze a provider export in AI FinOps, then read the same briefing built from the periods this browser keeps. Both are read in this tab.
Also on this site · the decision and release log
Shiplog records decisions, tracks the releases they shape, and keeps the link between them. A teammate who arrives six months late opens the release and reads the reasoning behind it, instead of reconstructing it from chat history. Recording, searching, filtering, exporting, and importing the log are all further down this page.
One recorded decision, and the release that shipped it
Representative synthetic records These invented records demonstrate Shiplog. They are not customer or internal operational data.
Try it now · nothing to install, nothing to upload
The Prompt coach scores one prompt for intent, efficiency, and model fit against a bundled rubric, names the single change worth making first, and lets you revise and grade again. A worked example is already graded when you arrive, so there is a real result to read before you paste anything. No sign-in, no upload, no storage, and no request to a model.
Open the prompt coachAlready have an export of your own AI history? Read your whole history instead — the same rubric across weeks of your own prompts, naming the one habit worth changing first and whether it is moving. Read in your browser tab; nothing is uploaded.
Keep reasoning with the work
Shiplog preserves what the team knew, what it considered, and who owned the call—then connects that reasoning to the release. New teammates get context without archaeology, and future decisions start with evidence instead of memory.
Background work was lost on deploys. The team compared a managed queue, database polling, and in-process retries before choosing a durable path.
The durable queue shipped alongside a read cache, keeping the release outcome connected to the decisions that shaped it.
Decision-making, unpacked
Get a concise field note on the habits, templates, and tradeoffs behind durable engineering decisions.
New record
Decisions and releases
Includes representative example records to demonstrate Shiplog. They are not customer or internal operational data.
Search updates as you type. Use ↑ and ↓ to move between records, and Home and End to jump. Press Enter to open details. Every row keeps its linked decisions or releases, whatever the filters hide.
Loading all decisions…
Portable history
Counting locally stored records…
The JSON file includes only decisions, releases, and their preserved links stored in this browser.
See what this browser is keeping — the retained records, how sure the page is about them, and how to erase them.
Portable history
Choose a file exported from Shiplog. It is read in this browser only — nothing is uploaded, and nothing is saved until you confirm the summary.
Representative release
This example uses the same release model and linked decision records as the full release history. Expand it to inspect the associated decisions.
Finding the latest release…