Gali Try Gali

Gali

Learn it. Review it. Out loud.

Talk through a PDF for free. Join the waitlist for full reviews and study tools.

Try it free Join the waitlist Watch the video
What holds up: claims Gali checked against the PDF and code. Code matches Algorithm 1, citing page 5 and dro.py lines 12 to 20. Batches follow the paper, citing page 4 and dro.py lines 26 to 37.

Review

Find out if a paper holds up.

Gali checks the math and runs the code while you read. Findings link to the page and line.

Watch a review
To study: practice problems and worked examples Gali made from the PDF. Practice: what does alpha do?, a question with its answer, citing page 5.

Study

Learn it by talking it through.

Ask Gali to explain a page or quiz you. It makes worked examples and practice problems.

Watch a study session

Waitlist only for now

Watch it work.

The Gali home on desktop: a drop zone that says Drop a PDF and its code, with a Choose files button. A file dragged over the home page. The whole page says Drop a PDF and its code. The CTC-DRO paper open on page 1. The Findings tab says No findings yet. The review has started. A live line in the Findings tab says Listing the main claims and the evidence for each, with a Stop button. Page 5 of the paper while the review keeps running. Page 5 with the sentence on smoother updates highlighted as note 1, while the review keeps running. The first finding, Smoothed updates behave as claimed, arrives while the review is still simulating CTC-DRO and group DRO. The finished review. What holds up: Smoothed updates behave as claimed, with a chart of the computed group weights under CTC-DRO and group DRO. The finding’s citations: p. 5 of the paper and dro.py lines 17 to 19. dro.py open next to the paper. Lines 17 to 19 are marked: the step divides the losses by the weight plus alpha. The paper on page 8, next to the finding’s p. 5 citation. The paper back on page 5. The cited sentence is marked, right below note 1. The paper with note 1 highlighted, next to the finished review. The browser shows ctc-dro-annotated.pdf, 1217.3 KB, downloaded next to the finished review. A study session on page 5 of the paper, with Equation 10 on screen. The Findings tab has no findings yet. Under To study: Worked example: one weight update, with a table of the computed weights. The worked example under To study, with its Add to notes button. Under To study: Practice: what does alpha do?, a question with its answer. The worked example’s Add to notes button. The Notes tab, Saved findings: the worked example, saved on page 5.

  1. Add a new paper.

    Drop the PDF, and its code if you have it.

  2. Run a full correctness review.

    Started the review. Keep reading while it runs.

  3. Highlight the sentence on smoother updates.

    Note 1, page 5.

  4. What did you find?

    It holds up. The numbers recompute and the code matches Algorithm 1.

  5. Show me the update in the code.

    Here, dro.py lines 17 to 19. It divides by the weight plus alpha, as in Equation 10.

  6. Where does the paper say that?

    Page 5, right below your note.

  7. Download my highlighted paper.

    Downloaded, with your highlights.

  8. Walk me through one update with real numbers.

    Added a worked example under To study.

  9. Now give me a practice question on alpha.

    Added one under To study, with the answer.

  10. Save the worked example to my notes.

    Saved to your notes, on page 5.

Where your data goes.

When you try it free, the PDF file stays in your browser. Gali sends its text and your questions to DeepSeek V4.1 Flash through Together AI. If you start a voice call, it runs through OpenAI. Your chat is not saved.

With full access, your projects live in a private sandbox that only you can open. Voice, typed chats, and what is on screen go through OpenAI. Answers about the PDF, reviews, and code runs use Kimi K3 and DeepSeek V4.1 Flash through Together AI. Gali saves your last conversation so it can remember it.

Gali is a research preview, so please don’t upload anything confidential. Read the Together AI Privacy Policy.

Questions

Do I need to sign up?

Not to try it: drop a PDF and ask Gali about it. Everything else needs full access for now: reviews, code runs, the To study list, saved notes, and your own private sandbox. To get it, join the waitlist. We email you when you're in.

What can I upload?

A PDF up to 100 MB. With full access, you can add code: a zip up to 100 MB or a GitHub link.

Does it catch everything?

No. Gali can be wrong. Our guide shows where AI review helps, where it fails, and how to check it. When you try it free, open the guide at the top of the page before you add a PDF. With full access, it is in the sidebar on a computer.

What can I take with me?

Your findings as text, your notes as Markdown, and the PDF with your highlights.

What does it cost?

Trying it is free, up to a limit. With full access, you get a funded trial, and your work stays when it ends.

Is Gali open source?

Yes. The code is on GitHub. Answers about the PDF, reviews, and code runs use open models through Together AI.