ILLATE

For teams with AI already in production

Cut your AI bill without hurting your answers.

I'm Poojith Devan. I measure where your OpenAI, Anthropic, Gemini or GPU spend goes on your real traffic, then cut it, and check every change against your quality bar before it ships.

20 minutes · no charge · no files needed. Not ready? Check your bill in a minute.

  • The aim: 20 to 40% off the bill, with every answer checked against your quality bar
  • First engagements from $750 (₹50,000), fixed price agreed before work starts
  • Baseline in 5 business days, finished in 2 weeks
00 See it measured

Teardown #1: a support bot, eight ways

One support workload, measured on self-hosted GPUs and priced on hosted APIs, with quality checked on 500 real questions. An independent demonstration, not client work.

  • About 40% lower API cost from prompt caching alone, on the same prompts
  • Twice the capacity on the same GPU from server settings, with no quality loss
  • One setting that cost 4 points of accuracy, caught by the quality check
  • Where self-hosting starts to pay: above about 1.7 requests a second, versus the cheapest APIs
01 Who it's for

Teams with an AI product in production

If one of these sounds like you, it's worth a 20-minute call.

Your API bill grows with every user

You call OpenAI, Anthropic or Gemini. I check whether prompt caching, shorter inputs and outputs, sending easy requests to a cheaper model, or batch pricing can cut the bill while meeting your quality requirements.

You run your own models

Your GPU bill keeps climbing, or you're deciding whether to self-host. I check GPU use, model precision and serving settings to find changes worth testing against your quality and latency requirements.

Responses are too slow

Users wait too long for the first word or the full answer, especially when traffic rises. I measure where that delay comes from and test what fixes it.

02 What an engagement can cover

We pick the parts that matter on the call

Each engagement has a defined scope. These are the areas it can include.

  • API cost: prompt caching, shorter inputs and outputs, and a lower-cost model where it's enough
  • Model routing: send easy requests to a cheaper model, checked against your quality bar
  • API or self-host: compare total running cost at your real traffic, including idle capacity, engineering time and reliability
  • Model efficiency: test lower precision (8-bit, 4-bit) to cut memory and cost, subject to quality and speed checks
  • Serving performance: find the bottlenecks and test vLLM, batching and cache settings
  • GPU sizing: compare suitable hardware against your throughput, latency and availability needs
03 How it works

Four steps, one fixed price

  1. A 20-minute call

    You tell me what you run, roughly what it costs and what hurts. No data needed. We find out whether there's a likely opportunity and what needs measuring. If there's no sensible next step, I'll say so.

    → a proposed next step, or no engagement
  2. Fixed scope and price

    One page: the workload, what I'll test, the quality, latency and traffic conditions every change must meet, the access I need, the dates and the price. First engagements start at $750 (₹50,000).

    → a signed one-page scope
  3. Measure, then change

    Within 5 business days of access you get a baseline: cost per successful conversation, latency and quality today. A cheaper model that fails more often isn't cheaper, so cost is always counted against answers that meet your bar. I then test changes under the same conditions and keep only those that meet the agreed bar. The report includes what didn't work.

    → before and after numbers
  4. Hand over

    Within 2 weeks you get the code and configuration changes, reproducible measurements, a rollout and rollback guide, and a handover call. Nothing reaches production without your approval.

    → changes your team can ship
04 After the call

What I'll need from you

Before any access, we agree where the work runs, what I can see, which tools may process it, and when access ends. I'll sign your NDA first. How I handle data →

  • Usage and billing summaries for a recent month
  • Your model and serving configuration, or which APIs and models you call
  • A representative set of requests to test with, or a safe way to generate one
  • The access agreed in the scope, granted through your own controls
  • A named technical contact on your side
  • Please don't email credentials or customer data
05 Questions

Questions people ask first

We only call OpenAI, Anthropic or Gemini. Is this still for us?

Yes, and that's where most of the easy savings are: prompt caching, a cheaper model for simple requests, batch pricing for work that can wait, and shorter outputs. Self-hosting only comes up if your traffic is high enough to pay for it. Try the AI bill calculator for a rough number.

What if there isn't much to save?

Then I'll tell you on the 20-minute call or in the written estimate, and there's no engagement. A paid engagement only starts once we've agreed there's a likely opportunity and what will be measured.

How do you make sure answers don't get worse?

Before changing anything, we agree a quality bar on a set of your real requests. Every change is measured against it, and only changes that meet it are kept. In Teardown #1, this check caught a setting that looked like the cheapest option but cut accuracy by 4 points.

Do you need our customer data or production access?

Not to start. The call and the estimate need no data at all. During an engagement I work on agreed samples inside your own infrastructure where possible, with access you grant and can remove. Details are on the security page.

Why work with one person instead of a firm?

You talk to the person who does the measuring and makes the changes, with no hand-offs. The scope is small and fixed, so you know exactly what you get and what it costs before anything starts. My background is in making models smaller and faster: published research on 1.58-bit language models and integer-only inference for chips.

How is it priced?

A fixed price, agreed in writing before work starts, based on the scope. First engagements start at $750 (₹50,000). The 20-minute call and the written estimate cost nothing.

Discuss your AI running costs

20 minutes with me, at no charge. No files or credentials needed. Choose a time and you get a Google Meet invitation.