import pandas as pd
df = pd.read_csv("designs.csv")
df["length"] = df["sequence"].str.len()
print(df[["name", "length"]])A team of AI scientists you lead.
Lumi Virtual Lab runs on your computer. Ask a research question, and a lead scientist answers it, runs the analysis, or convenes specialists and a critic. You see every step and every dollar, and you approve what matters.
Runs on your computer, with the model key you choose.
A short example. You ask which of three designs to test first. The lead scientist reads the file, convenes a team and a critic, then asks your permission before running a structure prediction tool.
From question to answer
The same chat handles a one-line question and a week-long investigation. The lead decides how much work it takes. See how it works.
You ask in plain language
Type a question, and attach a file if you have one.
You set the spending limit, and the screen shows what the chat has spent so far.
New chat$0.00 spentWe have three candidate nanobody designs for a new target. Which should we test first?
designs.csvThe lead gets to work
It looks something up, runs code, or answers from what it knows.
You can read each tool call and the code it ran, with the output. For a bigger question, the lead calls in specialists, or writes a plan and works through it as a project.
Lead scientist$0.07 spentinspect_data_filedesigns.csv has 3 rows: name, sequence, parent
name length 0 design_a 121 1 design_b 118 2 design_c 124Length alone won’t rank these. I’m convening the team.
The team challenges the answer
Specialists weigh in, and a critic pushes back.
In a team session, specialists take turns and the Scientific Critic challenges each proposal. You read the whole transcript.
Team sessionRound 1 of 3- Computational BiologistI want predicted structures for all three before we rank anything.
- Scientific CriticA predicted structure is a hypothesis, not a measurement. Say which assay would settle it.
- Lead scientistSummary: predict structures for all three, then rank them and show the caveats.
You get a report you can check
Every answer comes with how it was made.
The record shows which model spoke each turn, what it cost, and which tools it called. Export a chat or a project as HTML, PDF, or a Jupyter notebook.
Critic acceptedAntibody screen
Which nanobody design to test first
Design Fold confidence Recommendation design_b 86 Test first design_a 79 Test second design_c 58 Do not test yet Critic’s note. These are predicted structures, a hypothesis and not a measurement. A binding assay would settle the order.
Accepted in round 4$8.20 of $10.00 spent
You see what it does. You decide what it may spend.
An autonomous lab is only useful if you can trust the work. These limits are part of the product, not settings you have to remember.
- A hard spending limit
- Every project has a budget in dollars, and the screen shows what has been spent. Resume a stopped project and it does not pay twice for work it already finished.
- You approve the risky calls
- Orders to an outside lab and tool deployments stop and wait. You choose Allow or Decline, and the lab carries on either way.
- A critic that has to agree
- A project cannot finish until the Scientific Critic accepts the answer. If the critic objects, the team works another round.
- Code that runs in a sandbox
- Choose Docker, and generated code runs in a container with a read-only file system, no extra privileges, and limits on memory, processors, and time. It sees only the folder it works in, and gets none of your API keys unless you share one.
- A record of how each answer was made
- Every team session is saved with the model that spoke each turn, what it cost, which tools it called, and the software versions involved.
Antibody screen
Waiting for you- Round 1Team meetingDone
- Round 2Write codeDone
- Round 3Critic reviewObjected
Allow deploy_tool?
Proto will build this tool on your Modal account, which is billed for what runs.
What the lab can use
The lead picks the right tool for the question, and tells you which one it used.
- Public databases
Built-in lookups for protein, chemistry, and literature sources.
- UniProt
- Protein Data Bank
- AlphaFold
- PubChem
- ChEMBL
- Europe PMC
- PubMed
- arXiv
- An analysis toolkit
- More than 200 ready-made analysis functions, a catalog of over 100 bioinformatics programs, and a library of public datasets for the lead to draw on.
- Connected services
- Paperclip for literature search, Adaptyv for protein testing in a real lab, and Proto for structure prediction and design tools. These are separate companies with their own accounts and costs.
- Your choice of model
- Connect a key for one of nine model sources, including OpenRouter. The lab sends prompts to the provider you pick, so pick one that fits your data policy.
- Reports you can share
- Export a chat or a project as HTML, PDF, or a Jupyter notebook.
Yami, our protein foundation model, is in development.
Yami is designed to reason over sequence, structure, and physical plausibility together. It is not part of Lumi Virtual Lab yet. When it is ready, we plan to bring it into the lab.
In developmentQuestions
Something else? Write to founders@lumigenic.ai.
Does my data leave my computer?
The lab runs as a small server on your computer, and your chats, files, and results are stored in a folder you choose. Some things do go out: prompts go to the model provider you pick, lookups go to public services such as UniProt and PubMed, and requests go to any service you connect.
Lumigenic does not host the lab and does not receive your data.
Which models does it use?
You choose. Add a key for any of nine model sources, including OpenRouter, which reaches many models with one key. Keys are kept in your operating system’s keychain, and you pay the provider directly.
What does it cost to run?
You pay your model provider for what the lab uses. Every chat and project shows what it has spent, you set the spending limit, and a project stops when it reaches that limit.
Can it run experiments in a real lab?
Not on its own. It can connect to Adaptyv, a protein testing service, to order tests, and every order waits for your approval first.
Is Yami part of the lab?
Not yet. Yami is our protein foundation model, and it is still in development. When it is ready, we plan to bring it into the lab.
How do I try it?
Follow the steps on the Get started page to run it on your computer, or request a pilot and we will set it up with you.
Request a pilot
A pilot is a short trial of Lumi Virtual Lab on a problem you care about. Tell us what you want to investigate, and we will reply to set up a call.
Lumigenic is backed by 1517 Fund and has raised $100K.