A record your team can follow, a map of your process, the documents you were never trained to write, and an agent that learns your lab.
Everything below is produced from what you said in a session, in your words, and nothing leaves your workspace until you sign off. The five boxes are the short version; the rest of the page is the detail.
Your AI Transformation Assessment
The first thing your lab receives. Each step of your work placed on the ladder (captured, standardized, automated, transformed), what your lab already has and what it still needs, your top three priorities, and the first win to put to work, with today's baseline so you can see progress over time.
Your record
The protocol or process map: every step, input, output, tool and handoff, grouped into sections your team can own; every judgment call with its rule, or marked Not stated; the variance points where two runs could differ; the computational candidates. Exported to your notebook, repository or funder deposition when you approve.
Your process map and standardization map
The workflow drawn by role, generated from the record, and the map of each candidate: a plain rule or a model, how much of the step you still do, and how it ranks by how often it runs, how variable it is, how much it matters, whether it can be checked, and what a wrong result would cost.
Your template library
Twenty-odd documents, each filled from your sessions and rendered in your lab's layout and colors. Cards with a plain name, one line on what it is for, and who it is for. You add the ones that fit your strategy; nothing is pushed at you.
Your agent
It answers from your lab's own folders and cites the file each answer came from. It learns from every member's session and keeps what it learns in three memories. Every rule it keeps is visible with where it came from, and the PI confirms, corrects or retires it.
One team. One template. One standard.
You already have the science. What most labs never had is the operations toolkit that lets a whole team run on the same premise. These three come straight from the record.
Every session ends with a map: which steps already have a tool, which need a script, which values were never stated, which are ready to share. It is the first draft of your lab's documentation standard.
The protocol with the lab-specific values blanked becomes the template every member follows, so the postdoc, the technician and the rotation student produce the same record.
A script that processes results in a standard format, so every technician's findings run through the same analysis with the same settings, every time. You validate it before it touches real data.
The documents researchers were never trained to write, filled in for you.
Each card has a plain name and says who it is for. Open one to see it in your lab's layout; add it and it fills itself from your sessions, with anything that needs you marked Not stated. The standard business name appears once, small, so the person you hand it to recognizes it.
For my team
Who does what. Role descriptions. The protocol everyone fills the same way. How we document. Onboarding plan for a new member. What changed and who does what now. Our numbers.
For my PI or department
Program one-pager. The lab on one canvas. Where technology could help us. Cost of doing this by eye. Standard operating procedure. Project charter and milestones. What could go wrong.
For my funder
Your funder's data plan, in the format your funder requires this year. Budget line for computational support. Deposition-ready protocol. Progress since last time.
For a partner
Where we stand, only when you ask for an introduction. Compare suppliers, fee-free, only when you raised the cost.
An agent that understands your research and learns from your team's input.
The lens is not fixed at setup. Every member's session updates it: a new instrument, a new role, a value that moves from Not stated to stated, a correction you make at sign-off. What it learns is kept in three memories with different owners, and every rule it keeps is visible to you.
The science
What your lab studies and the words it uses. Used only to ask better questions, never to assert. Owned by the PI, confidential by construction.
The design
How your lab structures its programs: controls, acceptance criteria, the documentation standard it follows.
The procedure
Records, templates, scripts and roles. Shareable with a new member on day one.
It answers from your lab's own documents, and shows you where.
Point your agent at the folders your lab already keeps: protocols, SOPs, READMEs, analysis notes, the papers you work from. Methods indexes them and, when a question comes up, retrieves only the passages that bear on it (retrieval-augmented generation, or RAG). Every answer names the file it came from. Where a document and what you said disagree, it shows both and asks which one is what happens every run; that gap is often the most useful thing a session finds.
Cited, every time
Each answer points to the file and passage behind it. Nothing from a document is filled in where you said nothing.
Grows with the lab
When your PI shares the index with the lab, every member's folders join one index, so a new member starts from what the lab already knows.
Documentation, not raw data
It reads what you wrote about the work, not your instrument files. Raw data stays on your computer.
Every rule, with where it came from, for the PI to confirm.
When a session produces a new rule, the agent documents it into memory and highlights it for the PI or an authorized member. It is applied provisionally so the agent never stalls, and it stays highlighted until someone confirms, corrects or retires it. A lab can require confirmation before use. This is the screen, as designed.
| Rule | Where it came from | Status |
|---|---|---|
| Exclude a replicate when the Ct standard deviation exceeds 0.5 | postdoc · session 3 | Confirmed |
| GAPDH is the reference gene; stability not yet verified | PI · session 1 | Confirmed |
| Editing efficiency is read from the trace by eye | technician · session 4 | New · confirm? |
| Figures are regenerated from the table, not rebuilt in Prism | postdoc · session 5 | Corrected |
Why it is private
Nothing in any memory leaves your lab or trains a model. Each member owns their own record; the agent compiles across them only inside your workspace.
Why it compounds
A lab whose agent holds two hundred confirmed rules has a standard nobody else has and nobody can copy. That is the operating model, written by the people who do the work.
Your agent can tell you which tools already exist for your problem, and whether your own data could train something better.
Your agent starts with what your lab has written down. The next step is what your lab has measured. Most researchers do not know which tools already cover their problem, and almost none know whether the data they have already collected could power a model of their own. Ask, and your agent runs an advisory session that ends in a recommendation you can take to your team. It is measured, sourced and dated, and it says plainly what your data can and cannot support.
What exists today
The public tools for your problem, verified against primary sources with an "as of" date. Where a general model fails on your specific case, it says so.
What your data could do
Whether your own data could beat the public tools, and the honest boundary: what it does and does not answer. Type, rough count, how it was generated.
The full slate
Use public tools as they are. Fine-tune an open, commercially licensed model. Build a custom model from scratch. The case for and against each, so your team decides, not the software.
A phased path
Validate the signal cheaply on held-out data before anything scales. Licensing, data splits that prevent memorization, compute and on-premise privacy, and cost in people and compute.
Three options, the case for each, and where we would start.
An illustration for a lab that holds about a thousand proprietary structures and wants to predict where a small molecule binds. The rows are the slate; the last column is where each option sits in Methods.
| Option | The case for | The case against | Where it sits |
|---|---|---|---|
| Use the public tools as they are | No build, no license question, results this week | Trained on public structures; weak on the family your lab studies | Tier 1 · today |
| Fine-tune an open, commercially licensed model on your structures | Starts from a strong base; a thousand examples is enough to test the signal | License terms must be checked before anything commercial; base model choices change often | Tier 2 · the plan |
| Build a custom model from scratch | Entirely yours; nothing inherited | Needs far more data and compute than a thousand examples; the last option, not the first | Tier 3 · when validated |
Where we would start
Fine-tune first, on a held-out split, and measure against the public tools on structures the model has never seen. If the signal is real, scale with higher-value features. If it is not, you have spent a week, not a year, and you know.
What stays yours
Your data never enters Methods. Training runs on your institution's cloud or on-premise, under Tier 3, one approval at a time. The model finds and scores; you validate. The advisor never overstates, and it names what your data cannot answer.
What Methods could do next, and when it says so
Each skill listens during the session. It speaks at sign-off unless you chose, at the start of the session, to hear suggestions in the chat as you go. Either way every suggestion also lands in the documentation, and the session never becomes a sales call: nothing is priced, and nobody pays us to be suggested.
"Everyone in the lab does it a bit differently."
A standardized template for this protocol, with the values each person fills in and the rules that stay fixed.
"Our notebooks are all over the place." "The funder wants it deposited."
A lab documentation standard: one structure for internal records and external sharing, deposition-ready.
"I export to Excel and do it by hand." "Each tech runs the analysis their own way."
Automated intake: a script with a shared input form so every result is processed the same way.
"I copy the numbers into Prism and rebuild the plot."
A figure script that regenerates the plot from the table, with the axis and error-bar choices written down.
"I estimate editing efficiency by eye from the trace." "I draw the boxes in ImageJ."
An existing tool that already covers the step, named. Nobody pays us to be on that list.
"That kit is expensive." "We have always bought it from the same place."
A vendor comparison, fee-free, only when you raised the cost. Never a paid placement.
Six layers. Two are fixed and versioned. Four are composed for your lab, every session.
For the technical reader: how the session listens, what it compares against, what it offers, and why every record knows exactly what produced it. Read how it is built →