Skip to main content
The AI Cabinet

About

A cabinet. Not an agency. 

Two people who take on whole projects together. One works out what to build and whether it pays. The other makes sure it survives production. There is nobody else, and that is the point rather than the caveat.

Why two is the right number. 

We have both spent years on the other side of this: inside agencies, inside client programmes, watching the same failure repeat. A senior team wins the work, a junior team delivers it, and eighteen months later there is a platform nobody can operate and a maintenance bill nobody forecast.

What changed is that two people can now carry the workload that used to need a team, if they build with AI rather than talk about it. That is not a marketing line, it is the actual reason this works. It is also why the security discipline is not optional: throughput like that, pointed at production systems with real credentials, is exactly how organisations get hurt.

So the shape of the firm is deliberate. Two senior people, 3 active engagements at a time, no bench to keep busy, and a fixed quote before you commit. When a build genuinely outgrows us we say so and scope it separately, by name, in writing. We would rather lose the work than staff it badly.

We are generalists, not a sector practice. Between us we have shipped in real estate technology, coliving and property management, payments, insurance and consumer products. The skill we actually sell is learning an operation quickly and building for it honestly.

Who turns up

Two people. No bench. 

There is nobody to hand you to. That is the one thing an agency cannot promise.

What to build, and whether it pays

Mayank Pokharna

Eleven years in shared living and property operations. Brought in when the technology and the commercial reality stop agreeing. Builds with AI rather than talking about it, which is most of why two people can carry this.

Runs the workshops. Writes the spec. Tells you when you do not need us.

LinkedIn

Making it survive production

Atul Kumar

Fifteen years of engineering. A solutions architect who never stopped being hands-on: infrastructure, pipelines, deployment, and the security side, which in AI work is not a thing you add at the end.

Owns the architecture. Decides what an agent may touch. Answers the alert.

How we work

Six positions we actually hold. 

A firm with no position is a vendor. Disagree with any of these on the call and we will have a useful argument.

  1. 01

    Build first, then design, then fix

    The old chain of content, then design approval, then development is dead. We lock the specification, build to roughly eighty percent, put it in front of real users, and iterate. Rolling approval gates produce documents, not software.

  2. 02

    Produce the artifact, not a plan for the artifact

    Every session and every engagement ends with something you are holding. A working screen, a report, a running workflow. If a meeting produces only a decision about a future meeting, we got it wrong.

  3. 03

    The line you draw is respected

    Every client has something they will not hand to a machine. The front desk, the tenant relationship, the financial data, the hiring decision. We aim the work at the back office and leave the line where you put it.

  4. 04

    Fix the measurement before the workflow

    An automation you cannot measure is a story. If the tracking is not there we install it first, even when that is the less impressive month.

  5. 05

    We will tell you once, then execute

    If we think you are making a mistake we say so plainly, with the reasoning. If you restate your position, we build what you asked for and stop arguing. You know your business.

  6. 06

    Boring architecture, on purpose

    Version-pinned models, obvious data models, fewer moving parts. The interesting choice is usually the one that is hard to operate at 2am, and one of us has been the person awake at 2am.

The boring part

How we contract. 

Contracting terms
BillingMonthly in advance. USD via Wise or SWIFT.
Fixed-scope workThe Survey, Blueprint, Build and Hardening are each quoted as a fixed number before you commit. Not a range, and never an hourly rate.
Ongoing workMonth to month, 30 days notice either side, rate held for the engagement. No minimum term.
Not includedThird-party software and model API costs. These sit on your own accounts so you see them directly and they stay yours.
OwnershipYours from the first commit. Code, prompts, orchestration, eval sets, runbooks, documentation.
NDAStandard, and we will sign yours. We also have a mutual version that covers AI tool use, screen-share exposure and recording consent, which most templates miss.
Larger buildsWhen a build outgrows two people, there is a Delhi team we have worked with for years. Scoped, quoted and named in writing before anyone touches your systems. Never folded into our fee.
The checkpointOn any ongoing engagement, at four to six weeks, we ask one question: is this worth what you are paying? We give our own answer first.

Still reading? Let us talk. 

Thirty minutes. No deck. Sometimes the answer is that you do not need us, and that is a short call.

Book thirty minutes

3 engagements at a time