EVE:
AI assistant for researchers: computation, analysis, and code on demand
EVE writes the program your project needs, runs it on a server, has a second agent recheck every step, and delivers the code, tables, and figures to you in Telegram. You review, you decide, you remain the author.
Video demo with Marc, on a task of yours · 20 minutes · No commitment
- You
- Voice note · 0:31
- EVE
- Got it. 412 PDFs in the folder. I’m writing the extraction script (year, journal, country, sample size), running it on the server, and I’ll message you when it’s done.
- EVE
- Done. 406 articles in the table, 6 PDFs with no usable text, listed separately. A second agent re-read 40 articles at random: 2 sample sizes corrected, details attached. Script and README are in the folder.
- You
- Bootstrap confidence interval of the mean, 10,000 draws, fixed seed.
- EVE
- Running. I’ll be back with the table, the figure, and the script.
Example. We connect the shared folder with you, if your project needs one.
The principle
A frontier model, and the scaffolding that puts it to work.
The model does the reasoning and writes the code. A chatbot in a browser tab lacks everything else: memory, a server, a calendar, a check, and a channel where the files reach you. That’s what MACCUS installs around it, and that’s what holds it all together.
Memory
The project’s conventions, the exclusions decided in a meeting, the variable names, the decision made in May: it keeps them and still applies them six weeks later.
Server
It runs code on our servers. A script that takes an hour runs while you do something else.
Schedule
A literature search every Monday, a script rerun every night on new data: it notes the task and runs it at the set time, without you thinking about it.
Check
At every step, a second agent with no access to the first one’s reasoning redoes the computation from the files and reports the discrepancies.
Telegram
You type, or you leave it a voice note on the train. Long tasks run in the background and it messages you when they are done.
Files
Scripts, tables, figures, reports: it drops them in the conversation with a README, and in your shared folder if we connect one.
Job description
Its job, from the heaviest task to the most routine.
The first three tasks are the ones we set it up for here. The next ones revolve around the same project.
- Write the program, then run it
- Sort, clean, classify, match, extract: a dataset, a corpus, interview transcripts, hundreds of PDFs, a reference list. It writes the program for your exact case, runs it on the server, and delivers the result with the code. Deduplicating 12,000 references, recoding a questionnaire, pulling five fields out of 400 articles, splitting a corpus by speaker: that kind of work.
- Analyze and run the statistics
- Descriptives, tests, regressions, mixed models, survival analysis, whatever your question calls for. Every number traces back to a line of code you can read, change, and rerun. No table that came out of nowhere. Python by default, R if that’s your team’s language: we set that up during onboarding.
- Simulate, resample, permute
- Bootstrap, permutation tests, Monte Carlo, sensitivity analyses, parameter grids: tens of thousands of repetitions, at a scale nobody could reach by hand. You pick the random seed, it sets it, and with the same library versions the result reproduces exactly.
- Build a small tool for your method
- A converter between two formats, a consistency checker between two files, a minimal interface for a repetitive step of your protocol. It codes it, documents it, and you keep it.
- Search the literature and verify the references
- Deep research on a question, delivered as a report with its sources. Every DOI is verified against Crossref: it must resolve, and the title and first author must match. A reference it can’t verify is removed from the report.
- Track publications
- Alerts on your keywords and journals. It messages you when there’s something new, and stays quiet otherwise.
Besides the number crunching
The paperwork that eats into research time.
The rest of the job: tasks that need no scientific judgment and still eat up your evenings.
- Funding calls and deadlines
- The funding calls you follow, or ones it spots on your topics, the list of documents to submit, and reminders on the dates you set. It messages you when a deadline is coming up, and it nudges you again, on its own, about anything you asked it to track.
- Ethics applications, data management plans
- A first draft written from the project’s notes and what it has in memory: where the data is, who can access it, how long it’s kept. You mark it up, it revises the draft.
- Progress reports
- Your report to the funder or the lab, put together from what it has already logged: what was done, when, and with what results. You review it instead of reconstructing it.
- Conferences
- Abstract and registration deadlines, a first draft of the abstract from your results, travel options to compare, and the expense report prepared from the receipts you photograph.
- Inbox, calendar, meetings
- Inbox and calendar triage, if we connect them. Meeting notes from a two-minute voice memo, with what’s left to do, who does it, and the reminder to go with it.
- Teaching
- Exam dates logged with a reminder for each (or added to your calendar, if we connect it), the grading spreadsheet kept up to date from your files, replies to the student emails you pass on, drafted for your review.
Not in the job
It doesn’t sign your papers and isn’t their author. It doesn’t make phone calls. And nothing is connected by default: a Drive, a mailbox, or a calendar is added with you, with your credentials, and priced in the quote. For the rest of a company’s admin work, see EVE’s general job description.
On your machine
EVE on your own computer, as an option.
A small agent installed on your machine lets EVE work beside you, in its own window, while you keep working.
- Open and drive most of your apps in the background, and your browser in its own window, without taking over your mouse.
- Read, file, and sort the documents in your home folder.
- Use a program installed on that machine, including one licensed to that computer only, when it can be driven through its menus and buttons or from the Terminal.
- Use your reference manager and the university portals you’re already signed in to.
- Pick up a session you left open in the Mac’s Terminal app, a Claude Code session for instance, read its output, and type into it.
- Passwords, keys, keychains, and password managers: never touched. A text file in which it detects an access key or a token isn’t copied; you’re told where it is.
- A list of sites the agent can’t open: banks and payment services, webmail, account security pages.
- Every action is logged on your machine. The apps and pages it opens, and the files it copies, changes, moves, or deletes, are mirrored to you on Telegram as they happen.
- It takes over one of your windows or tabs only when you ask, and each takeover is announced to you on Telegram before it happens.
- A file it deletes goes to the Trash, where you can restore it.
- A kill switch, under your control.
- macOS only, today.
- Set up on request, never by default: we look at your machine and your software first, and it’s added to the quote.
- It only works while the computer is on and awake.
- A few steps can’t be done in the background, a save dialog for instance: it then asks you for the click.
- It works in your signed-in browser and with the access you already have on that machine, nothing more, minus the sites on the blocked list.
The method
One step, one check.
In research, a counting error costs months. So the rule is simple: nothing moves to the next step until an independent agent has rechecked the previous one.
EVE writes, runs, documents
The script, its run on the server, the result, and a log of what it did: files read, rows excluded and why, parameters, random seed.
A second agent redoes the computation
Fresh context, no access to the first agent’s reasoning. It starts from the files on disk, recounts, re-verifies every reference, and returns a verdict with its list of discrepancies. Until the check passes, the next step doesn’t start.
You review, you decide
You receive the result, the code, the log, and the check’s verdict. Everything written down can be discussed, redone, and cited in your methods.
This is the rule we apply to our own research work. Your review is still needed, and because everything is written down, you can do it from the record.
Readable and rerunnable
Code you can read and run again.
Every delivery includes the script, its parameters, the list of files read, and a README. You rerun it on your own machine, change a parameter, attach it to your submission if the journal asks for the code. Nothing is locked inside a tool you could not open.
Here is the kind of script it delivers for the request in the excerpt above: short, commented, with the seed fixed and the outputs named.
# ci_bootstrap.py
# Input: articles_table.csv (406 rows)
# Outputs: ci_bootstrap.csv, ci_bootstrap.png
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
rng = np.random.default_rng(seed=20261103)
table = pd.read_csv("articles_table.csv")
x = table["sample_size"].dropna().to_numpy()
boots = np.array([
rng.choice(x, size=x.size, replace=True).mean()
for _ in range(10_000)
])
lo, hi = np.percentile(boots, [2.5, 97.5])
print(f"n = {x.size}, mean = {x.mean():.1f}")
print(f"95% CI = [{lo:.1f}, {hi:.1f}]")
pd.DataFrame({"bootstrap_mean": boots}).to_csv(
"ci_bootstrap.csv", index=False
)
plt.hist(boots, bins=50)
plt.axvline(lo)
plt.axvline(hi)
plt.savefig("ci_bootstrap.png", dpi=150)Your data
Where your data goes, and who processes it.
An ethics committee or IRB application asks where the data is stored and who processes it. Here is how that works with us, including what we don’t guarantee.
- Your conversations, its memory, and your files are stored on our servers, in France.
- Your EVE has its own system account, its own network, and its own files, never shared with another client.
- Your messages travel through Telegram. The model is Anthropic’s, on your own Claude subscription. OpenAI transcribes your voice notes and indexes its memory, ElevenLabs produces its voice, both on our accounts. Anthropic, OpenAI, and ElevenLabs are US companies, so some of your conversations are processed in the United States.
- Connected to a Drive or a mailbox, it only has the access you open to it, and what it brings back goes through the same providers.
- A backup every night, the last 14 kept. If you cancel, everything is deleted at once, and the last backup disappears 14 days later, or immediately if you ask.
- We don’t open your conversations without a concrete reason. If we have to open one to fix an incident, we tell you.
By default, the model is Anthropic’s and part of the processing happens in the United States. For a project where that is not acceptable, EVE can be configured differently. Three options, decided project by project, with you:
- Claude in a European region of a cloud provider: Amazon Bedrock offers inference profiles confined to European Union regions (Frankfurt, Paris, Ireland, Stockholm, Milan, Spain), and Google Cloud a “Europe” multi-region endpoint limited to EU member states. Both state in their documentation that requests stay within that geography. The cloud account is in your name or your institution’s, and the model is billed there by usage.
- Open-weight models served from Paris: Scaleway hosts Mistral, Llama, Qwen, DeepSeek, Gemma, and others in a Paris data center, with zero data retention by default and no training on your data, under its published policy.
- Routing to EU-hosted providers through OpenRouter, on its business plans: requests go through a European endpoint that refuses to leave the region. Its definition of Europe includes the United Kingdom and Switzerland.
These three options cover the model. Voice-note transcription, memory indexing, and its voice go through OpenAI and ElevenLabs in the United States by default: we look with you at what can change there, and what can’t.
The trade-off: the newest models don’t always reach European regions on launch day, some are not offered there at all, and in our experience open-weight models still trail on the heaviest computation and code. So we choose with you, project by project, and we write down which provider sees what, voice notes included.
EVE works on what you give it. For interviews, the basic rule stays yours: pseudonymize before sending, keep the linking key out of its reach, and tell it what it must never write into its memory. The conversation itself is still stored on our servers, as stated above.
You can ask it at any time what it has kept from a project, have it correct a note, or delete one. The full list of our providers is in our privacy policy.
Several EVEs can share one server, each in its own account, network, and files. That is hosting separation, not a vault between two teams. If your protocol requires full separation, we put you on a server that hosts only you: a question to ask before, not after.
Integrity
You remain the author.
The byline stays yours
EVE prepares, computes, checks, and documents. The choice of method, the interpretation, and the signature are yours. Elsevier, for one, asks authors to disclose AI use in the manuscript, and to describe it in the Methods section when AI was part of the research itself. A tool can’t be listed as an author (Elsevier’s generative AI policies).
Easy to write, because everything is logged
Its log lists what it did, step by step: which script, on which files, with which parameters, checked by whom. An AI-use statement takes a few lines from there, with nothing to reconstruct from memory.
Verified, or dropped
A reference invented by an AI tool gets spotted on a reread, sometimes only after submission. Here, every DOI is verified against Crossref before it enters a report, and whatever cannot be verified is dropped. Preprints are cited as preprints.
Getting started
Three steps. Two are ours.
We look at your project
The data work you do by hand or keep putting off, your data formats, Python or R, where the data may go and where it may not. Its first name, its language (French, English, or Dutch), your time zone. That conversation decides the configuration.
We install it
We create its Telegram account and its space on our servers. We write its personality, note your project’s rules, and connect what was agreed. Then we check that everything runs before handing it over. Nothing to install on your side, no server to rent.
You give it a first file
You open the link we send and send it any message. You sign in once to your Claude account and send us the code it shows. Then you send it a dataset and a question.
The assistant itself is up within minutes. The connections to your tools and the European configuration depend on what we set up with you.
The terms
Requirements, price, and teams.
- Telegram, on your phone or your computer. Free, two minutes.
- A Claude subscription in your name, Pro or Max plan. EVE runs on your subscription, not ours: you see what you use and you don’t buy resold credits with a markup. We tell you which plan to pick. With the European configuration, the model is billed by usage on a cloud account instead.
- Your data in a readable format: CSV, Excel, PDF, text, images. It reads what you send it in Telegram, and what is in your Drive if we connect it.
Quoted after we talk.
A setup fee, paid once: the full setup, the bot’s creation, its personality, and your project’s rules.
Then a monthly subscription: hosting on our servers, updates, voice, memory, and email support.
Connections to your tools and the European configuration are priced separately. You pay for your own Claude subscription.
- One EVE per person: its memory, rules, and habits belong to one researcher.
- For a team, one EVE per member, each on its own Claude subscription. One quote covers all of them.
- A shared folder, connected with you, so scripts and results stay in one place.
Before you sign
The questions researchers ask us.
Can it work on interview data?
How do I know it won’t invent a reference?
Python or R?
What does it remember about a project?
Can several people in a lab use it?
What if I cancel?
Want to talk?
Bring a computation you keep putting off.
Twenty minutes on a video call. You describe a data task you do by hand, or never found the time to do, and we show you how EVE takes it on. We also tell you what we can connect on your side, and what we can’t.
Free · No commitment · No cold calls, ever
Reply within 24 hours.