How Lumi Virtual Lab works

You talk to one scientist, the lead. What happens next depends on the question: a quick answer, a team session, or a project that runs for hours.

Who is in the lab

The lead scientist
The one you talk to. It plans the work, decides when to bring in help, and answers for the team.
Specialists
Eight standing roles. In a project, the lead chooses who joins, or you can make the whole lab the team.
The Scientific Critic
Challenges the team’s proposals, and decides whether a project’s final answer holds up.

Three ways it works

The lead picks the one that fits. You can also ask for a project directly.

A chat

Most questions start here. The lead answers directly, looks something up in UniProt or PubMed, or writes and runs code and shows you the output. When a question needs more than one view, it consults a specialist or brings in the whole team.

A team session

The specialists take turns over several rounds. The lead sets the agenda, the critic challenges the proposals, and the lead closes with a summary. You can read the full transcript, along with a record of which model said what and what it cost.

A project

For work that takes many steps, you give the lead a goal and a budget. It plans, works in rounds, and stops when the critic accepts its answer or the budget runs out. The steps are below.

Inside a project

Every step is saved as it happens, so a project that stops can pick up where it left off.

  1. The lead writes a plan

    From your goal and your budget, it drafts a plan and chooses the team.

  2. It picks the next step

    Each round, the lead makes one move.

    It can hold a team meeting, ask one scientist, write and run code, change the team, or propose a final answer.

  3. Findings become notes

    What each step found is saved as a short note.

    Later steps get the notes that matter, chosen by the lead or by best match.

  4. The critic reviews the answer

    When the lead proposes to finish, the critic accepts or objects.

    An objection sends the team back for another round.

  5. It stops, and you decide what next

    A project ends when the critic accepts, when it reaches its budget or round limit, or when it stops making progress.

    Raise the limit and start it again, and it reads back the finished steps without paying for them twice.

Steering and approvals

While the lab works, you can pause it after the step that is running, resume it, stop it, or leave a note that the lead reads before its next step.

Some calls pause on their own and wait for you: any that orders from an outside lab, deploys a tool, or otherwise creates, submits, or pays for something. You see what the call is and why the lead wants it, then choose Allow or Decline. If you decline, the lead says what was declined and carries on without it.

Running code is a setting. Choose Docker for a sandbox, or let code run directly on your computer, which is not isolated. Code is off until you choose.

Give it your data

Your files
Upload a file or point to a folder on your computer, and describe it in a sentence. The lead reads the description to know when to use it. Files you point to stay where they are, and code can only read them.
Papers
Give the lab a set of papers from uploads, a folder, or bioRxiv. It reads them, finds the tasks, databases, and software they share, and can write functions for those tasks. You review the functions before any chat can use them.
Benchmarks
Score the lab on published question sets, including Humanity’s Last Exam and LAB-Bench, under a cost limit. Open any question to check the score against the answer.

Where it runs, and what leaves your computer

The lab is a small server on your computer. Only your computer can reach it, and it asks for a token. Chats, projects, files, and results are stored in a workspace folder you choose, and your model keys are kept in your operating system’s keychain.

Three things do leave your computer:

  • Prompts go to the model provider you choose, including any text the lead reads from your files and puts in a prompt.
  • Lookups go to public services such as UniProt and PubMed, and requests go to services you connect, such as Paperclip, Adaptyv, and Proto.
  • Code you allow to use the network can reach the network.

There is no Lumigenic cloud. We do not host the lab, and we do not receive your chats or your data.

See what it does with your question.

Tell us what you are investigating, and we will set up a call. Or run the lab yourself.