AI-powered analytics
There’s more in your data than your dashboard shows.
Reveal what’s happening, what’s driving it, and what to focus on next — while AI is handling data science behind the scenes.
Which customers are most likely to churn?
Customers at risk of churn
Tuklasin found that purchase frequency and days since the customer’s last purchase were the strongest factors associated with churn.
Moderate class imbalance — only 8.2% of customers in this dataset churned, so treat the model’s confidence with care.
What mattered most
- Purchase frequency
- Days since last purchase
- Support tickets in the last 90 days
- Average order value
Suggested next action
Investigate whether customers with declining purchase frequency become inactive before they officially churn.
Technical detailsExperiment #42
- Method
- Gradient-boosted trees (XGBoost)
- Validation
- 5-fold cross-validation
- ROC-AUC
- 0.927
- Compared against
- Logistic Regression · Random Forest
Illustration of a result screen. Not a live product.
How it works
Three steps, from a spreadsheet to a decision
The same path every time, so you always know where you are and what happens next.
Upload or connect data
CSV, Excel, or a SQL database. Tuklasin reads it the moment it lands — columns, types, missing values, and the problems worth knowing about before anyone acts on them.
Ask what you want
A business question in your own words, with no method to choose. Tuklasin also proposes investigations you would not have thought to ask for, each with the reason it came up.
Tuklasin analyzes and explains
You pick what runs. The analytical engines compute it with visible progress — the stage it is on, and how long is left — and the finding comes back in plain language, with its warnings and what to look at next.
What it reads
It understands the data before it proposes anything
This is what Tuklasin sees when you connect a dataset — and what it flags without being asked.
customers_v3.csv
1,842 rows · 24 columns
| Column | Type | Missing |
|---|---|---|
| customer_idAn ID, not something to learn from | identifier | 0% |
| churnedOnly 8.2% of customers churned — a small group to learn from | yes / no | 0% |
| last_invoice_dateRecorded after a customer churns, so using it would hand the answer to the model | date | 4.0% |
| purchase_frequency | number | 0% |
| avg_order_valueSome values are missing | number | 2.1% |
| support_tickets_90d | number | 0% |
The warning sits on the column, not in a report
A problem travels with the thing it describes. There is no separate tab where the caveats live and quietly get skipped.
The trap is caught before you act on it
A column recorded after the thing you are trying to predict will produce a beautiful-looking result and a useless one. It gets flagged while the data is being read, not after you have believed the answer.
Nothing has run yet
Reading the data is automatic and happens the moment it arrives. Deciding what to do about it is still yours.
AI recommends.
It reads your data and proposes the questions worth investigating, each with its reason.
You decide.
Nothing executes until you pick it. A recommendation is a proposal, never a result.
The system executes.
Dedicated analytical engines do the computing and show real progress while they work.
Principles
Rules the product holds itself to
Analytical tooling is easy to make confident and wrong. These are the constraints that keep it honest.
You don't need to speak data science
Which statistical test, which algorithm, which validation strategy — none of that is a question you have to answer. The default experience is business language; the technical detail is there when you want it, and out of the way when you don't.
The AI never does the math
The language model reads, recommends, and explains. Every statistic, metric, and prediction is computed by a dedicated analytical engine. A number the model produced by itself would be a bug, not a result.
Nothing runs until you choose it
Recommendations are proposals. The system does not quietly run everything it can think of and present you with the winner.
Every recommendation explains why
A suggestion offered without a reason is incomplete. You should be able to disagree with the reasoning, not just the result.
Results ship with their warnings
Class imbalance, missing data, weak statistical evidence, overfitting, data leakage. The caveat travels with the finding instead of living in a panel you can skip.
Every investigation is a reproducible record
What ran, on which data, with which method and validation, and what it found — kept so any result can be found again, compared, and re-run.
Who it's for
Business teams with data but not enough capacity to analyse it
In practice that means one person: whoever ends up owning the numbers. The marketer running the reports, the ops lead forecasting demand, the founder opening the spreadsheet on a Sunday. The answer is written to be read by the person who asked the question.
Sales
- Which customers are most likely to buy?
- Why are sales declining?
- Which products drive revenue?
Marketing
- Which campaigns perform best?
- Which customers respond to them?
- Where are conversions being lost?
Finance
- What is driving changes in profitability?
- Where are unusual financial patterns?
- What might revenue look like next month?
Operations
- Where are we losing efficiency?
- What is causing delays?
- How should we forecast demand?
Customer Success
- Which customers are at risk of leaving?
- What factors affect retention?
- Which customers need attention?
Questions it is built to answer
- Why are sales declining?
- Which customers are most likely to buy?
- Which campaigns perform best?
- What is driving changes in profitability?
- Which customers are at risk of leaving?
- Where are we losing efficiency?
- Are there unusual transactions?
- What might next month's demand look like?
- Did performance change after that event?
- What should I look at next?
Find out what your data already knows
Upload your data. Let Tuklasin understand it. Discover what matters, choose what to investigate, and find out what it means — while Tuklasin takes care of the data science.
FAQ
How the product actually works
Do I need to know statistics or machine learning?
No. You bring the business question; which method answers it is decided for you and explained in plain language. The technical detail — the method, the validation strategy, the metrics — is available whenever you want to inspect it, but you are never required to understand it to use the product.
Where do the numbers come from?
From analytical engines, never from the language model. The model decides what is worth investigating and explains what came back; the computing happens in dedicated statistical and machine-learning code. Every number can be traced to the run that produced it.
Does it run analyses automatically?
No. It reads your data automatically and recommends investigations automatically, but nothing executes until you select it. Recommending and running are deliberately separate steps.
Can I see how a result was produced?
Yes. Every investigation is stored as a reproducible record — the dataset and version, the question, the method, the validation strategy, the results and the warnings. Records can be compared against each other and re-run. Nothing is a black box you have to take on faith.
What happens while a long analysis is running?
You see where it actually is — the current stage, the analysis being run, cross-validation progress, results so far, and an estimate of the time remaining. An indeterminate spinner is not considered acceptable.
Can my whole team use it?
Not yet. A workspace is for one person today — there is no sharing, no second seat, and no roles. “Teams” describes who Tuklasin is built for and whose questions it answers, not how many people can be in it at once. Working together on the same data is a direction we are interested in, not something we are promising.
Can I disagree with what it suggests?
That is the intended use. Every recommendation states its reasoning so you can reject it on the merits, and you stay in control of what runs, what gets revisited, and what gets ignored.