LLM Cost & Routing Sprint
Your spend is growing faster than your usage.
Three weeks: find where the money goes, fix the largest lines, and leave the controls that stop it coming back.
3 weeks · $15K-20K · Signed off by CFO or CTO
$200K/yr
Cut from cloud and LLM spend across client platforms
$40K/day
Peak burn contained in a production incident, within 72 hours
1.8M+
Tool invocations running under policy validation
Who this is for
- You have agents or LLM features in production and the bill has outgrown the forecast.
- Nobody can tell you what a single task costs, or which agent is responsible for the spike.
- You are being asked for a number by finance and you do not have one.
And who it is not
- You are pre-production. There is nothing to attribute yet, and the fix is architectural — that is the Substrate Build, not this.
- Your spend is under about $5K/month. At that scale the calculator on this page is the right instrument — the sprint starts paying for itself from roughly that level up.
What you get
- Cost attribution per agent, per model, and per tool, from your own traces.
- A routing plan that moves each task class to the cheapest model that holds quality, with the measurements behind each move.
- Prompt caching and context re-injection fixed at the integration layer.
- Runtime spend caps and circuit breakers at the tool-call layer, keyed to task type.
- A written teardown with every finding priced in dollars, not severity levels.
How it runs
Week 1
Attribution
Instrument and measure before changing anything. Per-agent, per-model, and per-tool cost mapping against real traffic, plus the tasks that dominate the bill.
Week 2
Execution
The largest lines get fixed: routing by task class, prompt caching, batching, and the token-heavy integration patterns that repeat across most codebases.
Week 3
Controls and handoff
Spend caps and circuit breakers shipped into your deployment path, dashboards handed over, and the teardown walked through with your team.
Relevant work and public code
01
Production emergency
Projected $40K+/day burn contained within 72 hours. No recurrence reported during the engagement.
05
LLM cost reduction program
$200K/yr aggregate reduction in cloud and LLM spend.
Read the code first
Sources and sample deliverables
experimental · Reviewed 2026-09-22
Meridian scheduling implementation. Synthetic replay and component checks are documented. The real-vLLM benchmark is marked unverified in this revision; no fleet-level latency or savings claim follows from it.
author reported · Reviewed 2026-09-22
Anonymized client outcomes. Reported by Nick Cerutti; client identities and raw operational records are not public. $200K/year is aggregate savings across engagements. $40K/day was projected exposure; 72 hours describes that incident, not a response or resolution guarantee. Zero-known-incident statements describe observations, not proof of absence.
Inspect a deliverable
Executable arithmetic with workload assumptions, quality gates and unmeasured latency. Free Markdown download; no email required.
What it costs
$15K-20K
Fixed for the scope agreed on the call. Moves with the number of services and models in play, not with how much you spend.
Every engagement can start as a $2,500 paid pilot, credited in full against the full scope if you go ahead.
Every engagement and what it costsBefore you talk to anyone
Agent Burn Calculator
What your agents cost per day, and what a runaway loop would cost before anyone noticed. Free, runs in your browser, and nothing you type leaves it.
Open the agent burn calculatorQuestions people actually ask
How much do you usually find?
It depends entirely on where you started. The published figure is $200K/yr aggregate across a portfolio of client platforms — not a per-client promise. If the first week shows the savings will not cover the engagement, that gets said in week one, not week three.
Do you need production access?
Read access to traces and billing is enough for the attribution phase. Write access is only needed for the execution phase, and can stay with your team if you would rather ship the changes yourselves.
Will this slow anything down?
Routing changes are measured against quality before they ship. The open-source policy engine publishes component benchmarks. End-to-end enforcement overhead is measured on your workload before rollout; it is not inferred from a percentage of a typical LLM call.
What if the problem turns out to be architectural?
Then you get told that, with the evidence, and the sprint fee is credited against the larger piece of work if you want it done.
Can we just use the calculator and do it ourselves?
Yes — that is what it is there for, and if one caching fix closes the gap, you are done. Where teams usually bring us in is when the number is bigger than one fix: attributing spend across agents, models, and tools takes instrumentation most codebases do not have yet, and that is week one of the sprint.
Fifteen minutes is enough to work out whether this is the right engagement, or whether it is one of the others, or none of them.
Book a 15-min callScoping by email works too — send what you're building and the read-back comes with a recommendation on where to start — even when the right start is smaller than you expected.
The written teardown: the seven integration patterns that account for most of the spend in the codebases behind the $200K/yr figure, and the arithmetic for each.
Request by emailn@nickcerutti.com · Include the question or project you have in mind.