How this works

AI assistant for Dhavan’s published portfolio material. Answers can be incomplete or wrong, so do not submit confidential code or data.

Your questions go through Hugging Face, and a model run by DeepInfra writes the answers. Chats may be retained for about 30 days to operate, debug and improve the assistant; DynamoDB TTL deletion is approximate. Cal.com handles call bookings. Details are in the privacy policy.

Use the suggested questions or ask about published projects, credentials, and engagements.

Independent AI engineer for software companies

I make AI products your customers can rely on.

When your AI gives wrong answers, runs up surprise bills, or nobody can explain what it did, I find the cause, fix it, and hand it over to your team.

Dhavan Shah

Dhavan Shah. You work directly with me. The first 30 minute call is free, and if I can’t help, I say so before you pay anything.

Built production AI features for 20+ customer workflows at DevRev · MSc Data Science, University of Zurich · Top 0.5% on Kaggle, Google’s data science competition platform · Founder, KazenAI

Built production AI features for 20+ customer workflows at DevRev · MSc Data Science, University of Zurich · Top 0.5% on Kaggle, Google’s data science competition platform · Founder, KazenAI

How it works

Check it, fix it, keep it working.

Prices in USD. Remote. Call times overlap EU, UK and US working hours, and the booking page shows them in your time zone. Fixed scope and written deliverables, not open ended billing.

1 · Check Start here

Production Readiness Diagnostic

$5,000 fixed · 2 weeks

I look at what your AI actually did and what it cost, then give you a written report: is it ready, what’s blocking it, and what to do next. Read access only to start; nothing changes in production. Start a Pilot within 30 days of the readout, and 50% of the fee ($2,500) is credited against it.

2 · Fix

Production Pilot

Typically from $18,000

I build the agreed work from a written proposal with options. It starts narrow and is tested before it reaches customers. Your team keeps the code, tests and runbooks, so you don’t depend on me afterwards.

3 · Keep it working

Part time AI engineer

One day a week from $5,000 a month

I work as part of your team on whatever is slipping: a new workflow, a cost regression, a failure that came back. Two days a week from $10,000. Start after the Diagnostic or a Pilot, or straight away if you already know what you need. 3 month minimum, then month to month.

In two weeks, you know what’s stopping your launch.

The $5,000 Diagnostic ends with a written report that answers three questions. This is what one looks like, with example findings.

Production Readiness Diagnostic report

Example, not a client result

1 · Is it ready?

Not yet. Two things are blocking launch.

  • How often does it get things wrong?

    Blocking launch

    Nobody knows. It’s only spot checked.

  • Could you stop one customer’s runaway bill?

    Blocking launch

    No. Spend is only visible per month, in total.

  • Can you show what happened when it went wrong?

    Needs work

    Partly. Logs exist, but can’t be tied to a customer.

  • Does each user only see what they’re allowed to?

    In place

    Yes, and tests prove it.

2 · What’s blocking it?

  1. 1

    No measured failure rate

    Customers will find the failures before you do.

  2. 2

    No limit per customer

    A few accounts could wipe out your margin.

3 · What should you do next?

  1. A

    Fix the two blockers

    Add a test suite built from your real traffic, and spending limits.

  2. B

    Fix all three gaps

    Option A, plus a receipt for every decision.

Start a Pilot within 30 days of the readout, and 50% of the fee ($2,500) is credited against it. Read access only to start; nothing changes in production.

Diagnostic payment: 100% upfront, or 50% on start and 50% before the final readout. Pilot milestones are set in the proposal. Want only the KazenAI product, with your team doing the integration? kazenai.com

The problem

A demo only has to work once. Customers need it to work every day.

All four happened to real companies, and all are on the public record.

Wrong answers

Cursor’s support bot invented a policy, and users said they were cancelling.

It told a customer the product worked on one device per subscription. No such rule existed. How many people actually left isn’t known.

Source: CX Today, April 2025

Do you know how often yours gets it wrong?

A measured failure rate

Tests built from your real traffic run on every change, so you know how often it fails and whether a fix helped.

This is us: book a call

Accountability

A tribunal held Air Canada responsible for what its chatbot told a customer.

The chatbot gave wrong advice about bereavement fares. The airline argued it wasn’t liable, and the tribunal rejected that. The award was small, about $812 with interest and fees, but the airline still had to answer for it.

Sources: McCarthy Tétrault · Dentons, on Moffatt v. Air Canada, February 2024

When your AI gets it wrong, can you show what happened?

A receipt for every decision

Every AI request, allowed or stopped, is recorded with the customer behind it, so you can show what happened and why.

This is us: book a call

Access

A flaw in Asana’s AI connector could show one customer’s data to another.

Asana told BleepingComputer about 1,000 customers were affected. It was a logic flaw, not a hack, and the report shows no evidence the data was misused.

Source: BleepingComputer, June 2025

Can you prove each user only sees what they’re allowed to see?

Permissions enforced and tested

It only reads what each user may see. It cites its sources, or says “I don’t know” when the evidence isn’t there.

This is us: book a call

Cost

“We are currently losing money on OpenAI Pro subscriptions! People use it much more than we expected.”

Sam Altman, CEO of OpenAI, on its $200 a month plan, January 2025. Six months later, Anthropic announced weekly limits on its plans, saying “one user consumed tens of thousands in model usage on a $200 plan.”

Sources: ITPro · Anthropic’s statement, quoted by Simon Willison

Could one customer cost you more than they pay?

A spending limit per customer

Each supported AI request is checked against that customer’s limit before it runs. It starts in shadow mode, which only watches, so nothing is blocked until you trust it.

This is us: book a call

If one of these could happen to you, that’s where we start.

The first 30 minute call is free. Pick a time that suits you.

Book a free call

One outcome: AI you can put in front of customers.

Why me

You work with the person who built it.

I’m an independent AI engineer. I make AI features and agents your customers can rely on: a measured failure rate, a spending limit per customer, and a record of every decision.

Dhavan Shah

Track record

20+
customer workflows where I architected production LLM features at DevRev
~2,500
automated tests across the KazenAI product
MSc
Data Science, University of Zurich
Top 0.5%
Kaggle Competitions Expert, on Google’s data science competition platform

Before DevRev, I did applied AI work with AXA XL and Sulzer in Zurich.

How I handle failures. When a review found that some calls in KazenAI could skip the budget check, I fixed it, published a security advisory and withdrew all 16 affected versions from new installs. What happened.

KazenAI is the product I built from this work. Clients can use it; it’s never required. KazenAI is a product, not a consulting service. Under the engagement agreement, what I build for you is yours, KazenAI stays mine, and your code, systems and confidential data never go into it.

I’ve already built the tools for this.

I don’t start from zero. I built these for exactly these four risks, and they’re how the Pilot gets delivered. Built by me, not client case studies. Agent FinOps is KazenAI’s product; run receipts are part of it.

Supporting capabilities

See all technical work

Is this for you?

Best when customers are about to depend on it.

Strongest fit: an AI feature or agent your customers use, in production or close to it, with a technical owner.

Good fit

  • A launch, a large customer, or a pricing change depends on the agent behaving.
  • Model spend or failures are growing faster than your confidence in them.
  • You can share traces, request logs or billing exports under scope.

Not a fit

  • Idea stage demos with no real users yet.
  • Open ended staff augmentation without a defined outcome.
  • Guaranteed savings figures, compliance certification, or work outside agreed scope.
Book a free call

The first 30 minute call is free. No obligation until you approve the Diagnostic.

Next step

Book a free 30 minute call.

Pick a time that suits you. The booking page shows times in your time zone and asks two short questions, what you’re building and what’s blocking it, so we start on your problem. If I can’t help, I’ll say so.

Book a free call

Cal.com handles the booking.

Prefer email? dhavan@dhavanshah.com This call is for consulting. Looking for the KazenAI product? Book a margin review on kazenai.com.