Founded on an argument about evidence
Sciematics Insights is a data science and artificial intelligence practice established in 2024, part of the Acmez group, with offices in Roorkee and Mohali.
A reaction to the demo
The Acmez group kept being shown impressive machine learning demonstrations that quietly collapsed in production, and kept meeting clients who could not tell the difference in advance.
The pattern was consistent. A model evaluated on data related to its training set. A metric with no baseline to compare against. An accuracy figure with no interval, quoted to one decimal place as though that implied precision. And, almost always, no monitoring after launch, so nobody noticed when the world shifted underneath it.
None of this requires new science to fix. It requires the ordinary discipline of an experimental field: hold out your test data, state your baseline, report your uncertainty, and watch the thing after you ship it. Sciematics was founded to do exactly that, and to say no when the data will not permit it.
Two years, unpadded
We were founded in 2024. We will not dress that up as a legacy.
Sciematics is founded
Established inside the Acmez group to do data science with the evaluation discipline the field claims to have.
The first refusal
A client asked for a churn model. Their data had eleven months of history and a leaked target column. We said no, and wrote down why.
Mohali works opens
A second office in Punjab brings us closer to a deep pool of machine learning and platform engineers.
Model cards become mandatory
No model leaves this company without documented limitations, subgroup performance, and a monitoring plan.
Thirty one models in production
Every one of them monitored. Two were retired when the evidence said they had stopped earning their keep.
Slower than we could be
We turn down work we cannot evaluate honestly. That is the constraint, and it is deliberate.
Four commitments, each testable
Measure against a baseline
A model that cannot beat a well tuned simple rule has not earned deployment, however sophisticated it is.
Publish the uncertainty
Every metric on this site carries a confidence interval. A point estimate alone is a sales figure.
Guard the test set
It is opened once, at the end. Anything else is measuring your own memory.
Retire models that stop working
Two of ours have been switched off. Keeping a decaying model alive to protect a case study is a form of fraud.
Return the burden of proof to the model
A machine learning system should have to demonstrate that it is better than the thing it replaces, under conditions it has never seen. Until it does, it is a hypothesis with a budget.
Be the firm asked to check somebody else's model
The most flattering engagement we receive is an independent evaluation of a system we did not build. It means someone trusted us to tell them something they did not want to hear.
Head Office, Roorkee
#651/52 Ganga Enclave, Near Sainik Colony, Right Bank Canal Road,
Roorkee 247667, District Haridwar,
Uttarakhand, India
Works, Mohali
SCF 20, First Floor, Shaheed Bhagat Singh Market, Sector 125, Sunny Enclave, Kharar,
Mohali 140301,
Punjab, India
Send us your data problem. We will tell you whether it is one.
A working data scientist reads every enquiry. You will get an honest read on whether your data can support the model you have in mind, before anyone quotes you a number.