Fixed scope, fixed price, read-only measurement first. I do not resell capacity, so nothing in the audit earns me a commission.
Inference audit
5–10 working daysfrom $4,500
- Per-provider table: published vs measured price, TTFT, tokens/second.
- Service-commitment reality: what is promised, what is met, what refunds.
- Model-name traps: aliases and snapshots that are listed, priced, and never served.
- Credential hygiene: key scopes, IP allowlists, spend caps, rate limits, pinning rights.
- Leak paths: presigned URLs and image links your request hands to the supplier's fetcher.
- Cache opportunity: prompt and response caching, with the saving stated in dollars.
One decision table, a 90-minute readout, and a prioritised fix list with owners.
Routing & cost architecture
2–4 weeksfrom $9,000
- A model policy per task class, with a floor price and a quality reservation.
- Fallback, cap and pinning semantics written as code, not as a wiki page.
- Caching strategy: what to pin, what to key, what to clear.
- An evaluation set and a regression harness so a model swap stops being a gamble.
- Reporting in your unit of work: cost per resolved ticket, per report, per 1k words.
A routing policy running in your repository plus a cost-per-unit dashboard.
Advisory retainer
monthly · roughly 2 daysfrom $1,200 / month
- Watching provider prices and behaviour so a change is a decision, not a surprise.
- New-model comparisons against your own traffic, not a public benchmark.
- "Which model for this" answers with a number attached.
- Quarterly re-measurement: the audit either still holds or you get the work back.
- Second opinion on hires and architecture reviews.
A standing decision log; unused days roll into the next month once.