One-time engagement
Compute cost reduction pass
A one to two week pass over what your AI system costs to run, read per resolved task: where the money goes by model, by step and by retry, and what to change first, from model routing and caching to prompt and context size.
The bill is growing faster than usage. Find out why, per task.
- 1 to 2 weeks
- Priced on the call
- A cost baseline and a ranked plan
Is this for you?
The bill is growing faster than usage
Each new customer costs more than the last, and the invoice does not say why.
Nobody knows the cost of one task
You know the monthly total. You do not know what one resolved request costs, so you cannot price it.
Margin is the next conversation
A buyer, a board or an investor wants unit economics, and the AI line is the one nobody can explain.
Not this? The architecture review reads the whole system for scale, with cost as one of six checks; the production readiness audit scores cost per resolved task as one of six gates on a live agent.
What you get
- 01
The read
Each of the six places the money goes, with your numbers from your traces and your bill.
- 02
The cost baseline
What one resolved task costs today, by model, by step and by retry, so you know your unit economics.
- 03
The ranked plan
Every change ranked by what it saves and what it takes, with the expected saving against each.
Illustrative pages. Yours carry your system’s scores, numbers and fixes.
Ranked savings
- 1EffortWorth
- 2EffortWorth
- 3EffortWorth
- 4EffortWorth
- 5EffortWorth
Cost per resolved task
| Model | Step | Retries | Per task |
|---|---|---|---|
Where the money goes
Six places, read on your bill
- Cost per resolved task
- Model routing
- Caching
- Prompt and context size
- Retries and fallbacks
- The model mix
Where the money goes
Six places an AI bill grows faster than usage. I read each one on your traces and your bill and put a number on it.
01Cost per resolved task
I start with the only number that matters: what one finished task costs once retries, fallbacks and human follow-up are counted.
I measureDollars per resolved task, by model, step and retry
02Model routing
I check which model handles each step and whether a smaller one would pass the same eval for a fraction of the price.
I measureEach step, the model it uses and the cheapest that passes
03Caching
I find the questions that come back again and again and check whether a cached answer could serve them.
I measureThe share of requests a cache would serve
04Prompt and context size
I measure what goes into each call, the system prompt, the history, the retrieved documents, and how much of it the answer needed.
I measureTokens in per call, and what the answer used
05Retries and fallbacks
I count the retries, the fallbacks and the long contexts that quietly double a task’s cost, and find where the curve bends.
I measureThe retries and fallbacks per task, and what they cost
06The model mix
I weigh the mix of models and providers against your evals, so quality holds while the cost comes down.
I measureThe model mix, with the eval score and the cost of each
1 to 2 weeks, a cost baseline and a ranked plan
The two weeks
- Days 1 to 2
Read the bill
Access in place; the billing data, the traces and the prompts as they run collected.
- Days 3 to 7
Put a number on it
Cost per resolved task built by model, step and retry; each of the six places read on your numbers.
- Days 8 to 10
Rank and write
Every change ranked by what it saves and what it takes, the expected savings stated, a readout with your team.
The next step
Start with a
20-minute call.
Tell me what the system does and what the bill looks like. You leave knowing whether the pass fits, how long it would take and what it would cost.
Bring
- Roughly what the AI line costs a month
- What one task is, in a sentence
- Which models you use, if you know
Why me
Cost control is part of running AI in production, and I have done it at both of the AI companies whose technology I led. At TWINii I cut database load by about half and own the token economics: cost per resolved task, by model, by step and by retry. At Devnagri AI the LLM workflows ran across 22 languages with LLMOps behind them, where every token had to be watched.
The track record“Mandeep didn't just put in a patch, he helped redesign the system architecture to handle 5x the load without too much of extra server hardware.”
Jaspreet Singh OberoiHead of Engineering, DevnagriAsked on most first calls
The questions.
How long does it take?
One to two weeks from the day access is in place. One system with one bill is usually a week; several services or several providers usually two.
What access do you need?
Your billing data from each provider, a sample of traces and the prompts as they run. Nothing is changed in your system.
Do you make the changes?
No. The plan is ranked and written for your team, with the expected saving against each change. If you want me on the changes, I can join part-time after the pass.
What if the bill is already lean?
Then you have a cost baseline per resolved task that says so, which is the number your board or your buyer was asking for.
Do you sign NDAs?
Yes, yours or mine, before I see anything.
What does it cost?
It depends on the size of the system, so it is fixed on the first call, in writing, before anything starts.
How is this different from the architecture review?
The review reads the whole system for scale, with cost as one of six checks. The pass goes deep on cost alone, with a number on every change.