AI is quick to write analysis code that runs on the first try. Your job focuses on what it can’t see: the question being asked, the quality of the data, the leaks, and what a result really lets you conclude.
AI helps data scientists with tasks such as “Get your modeling code reviewed,” “Take a first look at a dataset,” and “Write the data preparation code.” Start with a single task, in Claude Code: allow 15 minutes the first time, and always review the result before you use it.
For each task: how it goes today, then with the tool. And the time it takes the first time.
Claude Code · “Plan” mode
TodayA score too good to be true, and nobody available to hunt for the data leak before Thursday’s presentation.
With Claude CodeA list of comments ranked by severity: leaks, data split, metrics, reproducibility.
Julius · A CSV file uploaded to the conversation
TodayBefore the first modeling idea, it takes a day of charts and counts to know what the file really contains.
With JuliusAn overview of the dataset, with the charts and the list of pitfalls to handle before modeling.
GitHub Copilot · Agent mode, in VS Code
TodayPreparation code is long and repetitive, and it’s where the errors that skew everything else creep in.
With GitHub CopilotA readable, tested preparation module, with no leak between training data and test data.
Also
Elicit · “Find papers”
A table of the studies that compare methods on your type of problem, with what is settled and what is still open.
SciSpace · “Chat with PDF”
Reading notes geared toward reproduction, with the paper’s gray areas.
Cursor · “Plan Mode”: Shift + Tab
An organized, tested module that reruns the experiment in one command and gives the same numbers.
Hugging Face · “Models” tab, then HuggingChat
Two or three candidate models and a grid to compare them on your own examples.
ChatGPT · “Chat” mode, with thinking turned on
The sample size, the test duration, and a decision rule written before you look at the data.
Claude · With your results files attached
A two-page memo that says what the model does, what it misses, and what that changes for the decision.
Claude Code · “Manual” mode
An up-to-date model card, drawn from the code, with the items to confirm flagged.
All 10 use cases on this page are in Nova, each with its steps and its prompt.
Try the interactive demoWhat to check
Your code is sent to Anthropic for processing. Open a folder that contains the code, not customer data, and use a Team or Enterprise workspace for company code.
Get your modeling code reviewedClaude Code · “Plan” modeThe prompt to copy
Review the modeling code in this repository ([folder or notebook]) as a demanding machine learning reviewer. Look for, in this order: data leaks (features known only after the event to predict, statistics computed before the split, duplicates between training and test), a split that doesn’t respect time or groups, a metric poorly suited to class imbalance, badly built cross-validation, a missing random seed, and anything that would keep the result from being reproduced. For each comment: the file, the line, the severity, and a suggested fix. Don’t change anything.
In Nova, “My path” starts from your level: one next step at a time.
The first moves, with free tools.
In your jobWrite the data preparation code
In your jobGet your modeling code reviewed
In your jobSurvey the research on a method
In your jobTake a first look at a dataset
For a team: everyone moves forward in their own job. You see the team’s progress, never one person’s activity.
It reads the whole repository, notebooks included, proposes a plan, writes the code, and runs the tests. You review every diff before you keep it.
Recommended planPro, $20 per month ($17 per month if billed yearly). For company code: Team, $25 per user per month ($20 if billed yearly).
In VS Code, it completes code as you type, and its agent mode writes the data preparation functions with their tests.
Recommended planFree, $0: 2,000 code suggestions per month. At a company: Business, $19 per user per month.
Find out what the research says about a method before you commit to it: a table, one row per paper, with the references.
Recommended planBasic, $0. Then: Pro, $49 per user per month, billed yearly ($588).
Reading a tough paper: you highlight a formula or a table, and it explains it, citing the passage.
Recommended planBasic, $0. Then: Premium, $20 per month, or $12 per month billed yearly.
The public library of open AI models and datasets: this is where you look for a pretrained model before starting from scratch.
Recommended planFree account, with a small monthly usage credit. Then: PRO, $9 per month.
Outside the code: explaining a result to management, discussing an evaluation protocol, delivering a memo as a Word file.
Recommended planFree to start. Pro, $20 per month ($17 per month if billed yearly): it’s the same subscription as Claude Code.
Prices taken from the vendors’ official pages, dated on each tool page. These subscriptions are not included in Nova. Brands are named to identify the tools. No affiliation.
Nova recommends 6 to start with: Claude Code, GitHub Copilot, Elicit, SciSpace, Hugging Face, and Claude. 5 of them can be used for free to get started.
It reads the whole repository, notebooks included, proposes a plan, writes the code, and runs the tests. You review every diff before you keep it. The plan to choose: Pro, $20 per month ($17 per month if billed yearly). For company code: Team, $25 per user per month ($20 if billed yearly).
With “Get your modeling code reviewed,” in Claude Code (“Plan” mode). Allow 15 minutes the first time. A list of comments ranked by severity: leaks, data split, metrics, reproducibility.
No. Nova is software, not an accredited training provider: it issues no diploma or certification. It costs €19 per month for one person, and €129 per month for a team of up to ten people.