We would rather be useful than impressive.

Blutrain Private Limited is a small applied AI engineering firm in Zirakpur. We build production machine learning systems for organisations that have a real operational problem and limited patience for theatre.

01 / Why we exist

The industry has an honesty problem.

An enormous amount of money has been spent over the last few years on AI projects that never reached a single real user. Not because the technology was inadequate — it usually was not — but because the project was scoped by people selling it rather than people who would have to operate it.

We started Blutrain Private Limited because we kept being called in to rescue those projects. The pattern was almost always the same: an impressive prototype, no evaluation set, no baseline to compare against, no monitoring, no plan for what happens when the model is confidently wrong, and a data pipeline held together by a script on someone's laptop.

None of that is exotic engineering. It is ordinary, careful software work applied to a domain where a lot of people have decided ordinary careful software work is optional. It is not.

02 / Principles

Five positions we hold, including the inconvenient ones.

These are not values-page decoration. Each one has cost us work, and we would make the same call again.

01

A baseline, or the number means nothing

We will not quote you an accuracy figure without telling you what the current process achieves. '94% accurate' is meaningless if the rule you already have manages 93%. Establishing that comparison is the first thing we do and the last thing most vendors do.
02

The boring answer is often correct

A well-indexed database, a cleaned-up form, a scheduled report, a rules engine. We have closed several diagnostics by recommending exactly that, and billed only for the diagnostic. We would rather lose a project than sell you a model you do not need.
03

Errors have asymmetric costs

Missing a fraudulent transaction and flagging a legitimate one are not equivalent, and no default threshold knows that. We work out the real cost of each error type with the people who absorb it, before we tune anything.
04

Your team has to be able to maintain it

If a system can only be operated by us, we have built you a dependency, not an asset. We write documentation for the engineer who inherits the repository, and we consider an engagement failed if you cannot change the system without calling us.
05

Data protection is a design constraint, not a checkbox

Where personal data is involved we decide early what is collected, where it lives, how long it is kept and who can reach it — including whether it may leave the country or reach a third-party model provider at all. Retrofitting this is expensive and usually incomplete.

03 / What we hand over

Five layers, all of them documented.

A finished engagement is not a model file. It is a system with an owner, a test suite, a runbook and a cost profile. Every layer below is built, reviewed and transferred to your team.

Every layer we build, own and hand over documented

04 / The firm

Small, senior, and based here.

Where we are: Our office is at SCF 08, 2nd Floor, The Eminence Plaza, Ambala–Chandigarh Expressway, Zirakpur, Punjab 140603, India. Being in the Chandigarh–Mohali–Zirakpur belt means we can be on a client's factory floor or in their office the same morning, which matters more for this work than it sounds.
How we staff: Engagements are run by a small senior team rather than a pyramid. The engineers in your scoping conversation are the engineers who write the code. This limits how many projects we take at once, deliberately.
Who we work with: Operations, risk, engineering and product leaders — usually people who have inherited a process that does not scale, or who have been handed an AI mandate and want a technically honest read on it.
How we start: Almost always with a paid diagnostic. It is scoped, time-boxed and produces a written assessment you own, whether or not you continue with us.
Languages: We work in English, Hindi and Punjabi. For voice and language work this is a practical advantage, not a courtesy — the acoustic and code-switching problems in Indian contact-centre audio are not well served by models tuned on US English.

Next step

Ask us something we might have to answer awkwardly.

We would rather have a hard technical conversation now than a difficult commercial one in six months. Bring the constraints, the messy data and the internal scepticism.

Get in touch Read the FAQ