Eight fixes we ship
Ranked by money in your traces, not by fashion. Each one ships behind your evals.
Symptoms in your traces
Each one maps to a fix below. Fix them before finance asks about the invoice.
- Cache hit rate under 50% on repeated prompts.
- Input tokens more than 10 times the output tokens.
- The same tool called 3 times or more in one run.
- One frontier model on every route.
- No cost per tenant, feature or request.
- 01Prompt caching
We order each prompt so its stable part comes first and hits the cache. Anthropic charges at least 90% less for cached input.
Caching - 02Context engineering
We cut system prompts, history and retrieved chunks to what the task uses. Fewer tokens, and often better answers.
Context engineering - 03Model routing
A router sends each request to the cheapest model that passes its eval. The frontier model gets only the hard cases.
Model routing - 04System-1 models
Classification, extraction and routing do not need text generation. Single-pass models such as Jev or Clairic return a typed value with a confidence.
System-1 models - 05Tool-call inspection
We read the trace of every agent run. Loops, retries, duplicate calls and oversized results become a ranked list of fixes.
Tool-call inspection - 06Tool re-engineering
Fewer, coarser tools. Filtered and paginated results. Errors the agent can act on. One call can replace several.
Tools for agents - 07Cost control
Cost per request, per customer and per feature, in your own dashboards. Budgets, rate limits and alerts per tenant.
Cost control - 08Model evaluation
An eval set built from your real traffic. Each new model runs against it. We switch when cost drops and scores hold.
Model evaluation
Public evidence
- Anthropic charges at least 90% less for input read from its cache. Anthropic: prompt caching
- Cloudflare moved one extraction step to a small model. That pipeline cost 49% less and ran 3.2 times faster. Cloudflare: Clairic
What could your AI bill be?
Four questions. The estimate shows before we ask for your email.
Your AI bill, before and after
An estimate. Change any answer.
1/4
Evals first. Then the invoice.
No change ships on a hunch. Each one must hold your eval scores before it reaches production.
- BaselineCost per request, and an eval set from your real traffic. Both sides sign them.
- AuditTraces, prompts, tool calls and the invoice. Every leak ranked by money.
- Ship behind evalsEach change runs against the evals first. A regression blocks the release.
- MonitorCost per request in your dashboards, budgets per tenant, and a monthly run of new models.
A change that lowers an eval score does not ship, whatever it saves.
What we need: read access to usage logs and traces, the prompts and tool code, and one engineer for 2 to 4 hours a week.
Three ways to work with us.
The same three ways as every Heinrichs.ai project. Here, the gain is the money your AI bill no longer costs.
Not sure which one? Check your AI bill: 4 questions →
It starts with the diagnostic week. It needs a saving large enough to share.
Paid from the gain
20% of what you save or earn in the first year. Then it is all yours.
We measure the gain in your own data, with a method we both sign.
You pay the 20% after we measure the saving, never more than the maximum in the proposal. Before that, you pay only the diagnostic week: if we measure no saving, the fixes cost you nothing. If the saving is less than 5 times the diagnostic week, the diagnostic becomes credit for other work.
What the diagnostic week gives you
- Your numbers today, and how we will measure them
- Every leak, scored and ranked
- A cautious estimate of the saving
- A proposal with the maximum you pay and the weeks of work
- Measured in your dataCost per answer, before and after, at the same quality score.
- Our gain guaranteeThe 20% follows the gain we measure, so a smaller gain costs you less. We also work up to 4 more weeks at no charge to close the gap. This holds if your team did its part: data and access on time, one person 2 to 4 hours a week, and the new process in use.
Diagnostic and solution in one week, for one fixed fee. The fee follows the gain.
- Before the weekWe agree the scope and the fixed fee.
- During the weekWe find the leaks and ship the fixes with your team.
- At deliveryThe fixes run on your real traffic. You keep the fixes and the week's report.
Work beyond the agreed scope needs your written approval. If the week finds a larger saving, it can become a project paid from the saving.
Open scope, in sprints of 1 or 2 weeks. You set the order.
- Before you startTell us what your AI does. We send the rate and an hour estimate.
- Each sprint1 or 2 weeks on the tasks you put first.
- After each sprintYou see the work and set the next order.
More terms
- Running costsYou keep paying your AI providers directly, at their prices. We take no share of your usage.
- What you keepYour license is permanent once the work is paid. If we stop operating, the code and your data pass to you.
- The next process, mapped at no chargeAfter the bill, the work itself: we map the processes your AI could take over, at no charge.
- If we disagree on a numberAn independent accountant, named in the contract, settles it with the signed method.
Your data and the evidence
- Your dataAn NDA and a data processing agreement (LGPD, GDPR) before we read any log. Read-only access. No AI training on your data. Access removed and copies deleted at the end.
- No lock-inAny provider, any model, any cloud. Your accounts stay yours, and so does every change we make.
How much can we save on our AI bill?
It depends on how your AI works today. The check above gives a cautious estimate in four questions. The diagnostic week measures it in your own usage.
Will the answers get worse?
No. First we build a quality test from your real work. A change that lowers the score does not ship, whatever it saves.
What is prompt caching?
Most requests repeat the same instructions and documents. The provider keeps that part ready and charges much less to read it again. Anthropic charges at least 90% less for cached input.
What is model routing?
A router sends each request to the cheapest model that does it well. Simple requests go to small models. Only the hard ones go to the largest model.
What is a System-1 model?
A small model that makes one quick decision in a single pass: a category, a number, or yes or no. It does not write text, so it is fast and cheap. Jev and Clairic are examples.
Do we have to change AI provider?
No. We work with OpenAI, Anthropic, Google, open models and any other provider. You keep your accounts and pay them directly.
What access do you need?
Read-only access to your AI usage and, for agents, their traces. An NDA and a data processing agreement come first. We never train AI on your data.
How long does it take?
One diagnostic week, then 1 to 4 weeks of fixes. The first fixes can go live in about 2 weeks.
We already have an LLM gateway and dashboards. Why you?
Good, we use them. A dashboard shows the cost. We change the prompts, routes and tools that create it, behind your evals.
Send us one month of your AI bill.
Send the invoice and tell us what your agents do. We reply with the first leaks we would chase and what the fixes cost.
Email us your AI bill Book a 20-minute call
From São Paulo, mostly remote, for Brazil and abroad. On request we visit during the diagnostic; travel outside São Paulo is billed at cost.
Who you talk to
Samuel HeinrichsFounder, São Paulo
Has built software since 2011 and led AI work since 2017. Postgraduate degree in Artificial Intelligence. Works in Portuguese and English.
We are taking our first clients and start at most 4 new projects a month. We reply within one business day, in Portuguese or English.