AI research papers
AI Agentic Assistants that research and write papers with you
You pick the problem and the benchmark, so you lead the research. The Scientist assistant finds where the current best method falls short, ranks ideas, and designs the experiments. The Researcher reads the literature and keeps only references it found and checked. The Coder reproduces the baseline and runs each idea. The Copywriter turns the results into a paper in your voice. You are the author, and nothing is submitted, sent, or paid for without your go.
From question to paper
A paper draft for you to review
- Your problem and the current best method
- Ideas, experiments, and ablations, all checked
- A paper you review and submit
Nothing is sent or shipped unless you asked for it or approved it, and money always asks.
Numbers in the paper come from the results the experiments saved, and a reference is listed only once it was found and checked. Money, sending, publishing, and submitting wait for your approval.
A closer look
An idea goes in.
Real work comes out.
A paper draft for you to review.
Illustrated workflow
01 / Set a goal
Set a goal
You say what done looks like: a ticket, a client, a deadline, or a draft.
How it works
From a goal to a result.
- 01
Set a goal
You say what done looks like: a ticket, a client, a deadline, or a draft.
- 02
The team works
The assistants take the goal, use the ready prompts and skills, and do the work in the tools you connected.
- 03
You approve the risky step
Anything that sends, publishes, ships, or charges waits for your Approve. Approve a kind of step once and that assistant remembers it, but money always asks.
- 04
You get the result
Checked experiments and a LaTeX draft with its tables, references, and a short AI disclosure for you to edit, ready for your review.
Example
Send the draft paper to your co-author
It sends to people, so it asks first.
The team
You, and the assistants.
You, the CEO
You pick the problem and the benchmark, approve compute that costs money, and are the author who decides what is submitted.
Scientist assistant
Leads the research. Finds the limits of the current best method, proposes and ranks ideas, designs the experiments and the ablations, and checks every number and claim against the results files.
Researcher (Scientist assistant)
The Scientist assistant in its research role: searches the literature in arXiv, OpenAlex, INSPIRE-HEP, and Europe PMC, and keeps only references it found and checked. A gap is marked as needing a source.
Machine learning engineer (Coder)
Reproduces the baseline, implements each idea, and runs pinned, seeded experiments on its own computer. A heavy training run goes on a GPU you control, and it starts after your Approve.
Academic writer (Copywriter)
Turns the results into a LaTeX paper, an abstract, and the reply to reviewers, in your voice. You edit, and nothing is submitted without your Approve.
Features
How the paper gets made
Numbers from the results files
Tables and figures are made by scripts from the results the experiments saved, never typed from memory, so a number in the text can be traced to a file.
References that were found and checked
A reference is listed only after the assistant found the paper and checked its title, authors, year, and identifier. A gap is marked as needing a source.
Reproducible code
Experiments run with pinned versions and fixed seeds, and the scripts are kept next to the results, so anyone can run them again.
A reviewer's read before you submit
Ask the Scientist assistant to read the draft as a reviewer would. It checks each claim against the results and lists the weak points and the experiments a reviewer would ask for.
A full research run on its own
Coming soon.
Integrations
The tools it starts with.
- Science databases
- GitHub
- Hugging Face
- Google Drive
- Weights & Biases (MCP server)
Built in
What your assistants can use.
Its own computer
Each assistant works on its own computer in StellarFirm cloud, set up with the tools its job needs, and hands you the files.
Roles you make yourself
Name a role we don't ship, say what it does and give it any of our skills or your own. Comes with Ultra and enterprise.
Long jobs in the background
A big job starts at once and runs in pieces in the background, streaming back as it goes. You can stop it any time, and sends and spending still ask.
Research tools
Web, news and scholar searches, scholarly works, video transcripts and market data, all read-only.
LaTeX documents you can edit
A paper or lesson sheet in LaTeX opens beside the chat with a preview, and you can edit the source and save a new version.
A running app beside the chat
The Coder starts the app it built and shows it live next to the chat, so you can click through it yourself.
Keep going until it is done
Say what done looks like and the Coder keeps going, up to ten more rounds, until the goal is met. /stop ends it.
Follows up on its pull requests
Switch it on for a Coder and it takes on failing checks, merge conflicts and review comments on its own pull requests. Merges and replies still ask you.
Ready to use
Prompts and skills.
Skills
- Research brief
- Literature search
- Generate and verify
- Reproducible analysis
- Research paper in LaTeX
- Edit and proofread
Ready prompts
Reproduce the baseline of this paper from its repository on this benchmark, and tell me how close the numbers are to the published ones.
List the limitations of the current best method on this benchmark, then propose three ideas and rank them by how much they might gain and how cheap they are to test.
Run an ablation of each component of my method with three seeds, and show the results with intervals.
Draft the paper in LaTeX with every table generated from the results files, and flag any claim the results do not support.
Check every reference in this bibliography: does the paper exist, and do the title, authors, and year match?
Find the closest prior work to this idea and tell me plainly if it has been done.
Write the reply to reviewers and run the extra experiments they asked for, for my review.
Questions
Before you join.
Start with your team.
Create your account, name your company, and your team of assistants is hired and ready, with a guide for the first task.