What a one-week audit looks like.

Week 1, by day

1. Day 1

   Intake call · access setup

   • · 60-min kickoff. I learn your stack, your call mix, your constraints.
   • · Mutual NDA signed if you want one (template ready).
   • · Read-only API access to OpenAI / Anthropic / Vertex / wherever your LLM spend lives. CSV export works too.
   • · You hand over ~50 representative prompts (anonymized is fine).

2. Day 2

   Spend pull · call-type classification

   • · Pull last 30–90 days of spend from each provider dashboard.
   • · Sample ~1,000 representative requests from your traffic.
   • · Classify each into one of seven task types (classify, extract, translate, brand_voice, vision, grounded, embed) via Claude Sonnet 4.6.
   • · Output: spend breakdown table + call-type mix percentage + concentration risk number.

3. Day 3

   Routing analysis

   • · Map each task type to current provider/model and to candidate alternatives.
   • · Per-task cost-per-1M-token deltas across Groq, Cerebras, Vertex Gemini Flash, Anthropic, OpenAI.
   • · Identify the top three routing changes by dollar impact.
   • · Output: per-call-type analysis (the §3 of your final report).

4. Day 4

   Quality vetting

   • · Side-by-side benchmarks on your sample prompts. Old provider vs proposed provider, scored.
   • · Reject any swap that regresses on accuracy, latency tail, or brand tone.
   • · Brand-visible writing stays on Anthropic Sonnet by default — that's a hard rule, not a recommendation.
   • · Output: routing recommendation table with quality-confidence column.

5. Day 5

   Code drop · fallback chain

   • · Drop in the routeByTask() entry point and 5 typed provider clients.
   • · Retry-with-jittered-backoff, 30s timeout, graceful degradation when primary is rate-limited or down.
   • · Telemetry hooks: per-task counter, latency histogram, cost-USD counter.
   • · Output: implementation-snippets/ directory + README ready to drop into your repo.

6. Day 6

   ROI · risk · report

   • · ROI calculator: payback weeks, year-1 net savings, year-2 net savings.
   • · Risk + mitigation matrix (drift, rate-limit, brand-voice, bill shock, policy change).
   • · 30/60/90-day rollout: shadow mode → 50/50 → 100% with safety nets.
   • · Output: report draft (PDF + Markdown).

7. Day 7

   Delivery walkthrough

   • · 30-minute call with your engineering team. I walk the report and the code line-by-line.
   • · Open-question Q&A. If you want the $15K turnkey implementation, we discuss scope here.
   • · One week of email Q&A starts after delivery — no clock-watching, no hourly fees.

What we need from you

• · Read-only API keys (or CSV exports) from each LLM provider you use.
• · ~50 representative prompts. Anonymize whatever you want — I don't need PII.
• · 30 minutes of engineering time across the week to answer clarifying questions.
• · A 30-min calendar slot on Day 7 for the delivery walkthrough.

That's it. No multi-stakeholder workshops, no questionnaires, no 50-page intake form. Most clients spend less than two hours total on their side of the audit.