Grounded answers
Retrieval with citations
Every claim links to a passage in a paper, a patent or a row in your results. If there is no source, Mentra says so.
Product 1 of 5 · Research copilot
Mentra reads the literature, patents and your own notebooks, explains why a candidate ranked where it did, and drafts the next campaign. It never launches work without a named person approving it.
Abstract
01 · Capabilities
Grounded answers
Every claim links to a passage in a paper, a patent or a row in your results. If there is no source, Mentra says so.
Explain a ranking
Breaks a ranking into hull distance, predicted property, uncertainty and the filters that applied.
Draft a campaign
Turns a question into a campaign file with search space, filters and a GPU budget for you to edit.
Approval-gated actions
Proposes screens or lab batches; a named person approves each one in the console.
Your archive
Indexes past DFT runs, reports and lab notes in your tenant, so earlier work is not repeated.
Guardrails
Keeps answers inside the documents a user is allowed to see, and logs every retrieval.
02 · Method
A question in plain language, scoped to a program.
Search literature, patents and tenant data; rank passages by relevance and recency.
Compose an answer only from retrieved passages; mark gaps explicitly.
Draft a campaign or batch with its GPU cost for approval.
03 · Specification
| Item | Detail |
|---|---|
| Inputs | Questions; literature and patent corpora (licensing per source); tenant documents and results |
| Outputs | Cited answers, ranking explanations, draft campaign files, approval requests |
| Actions | Start a Dynara screen, propose an Inventa batch, route to Synthex: all approval-gated |
| Models | NeMo-based LLM with retrieval and guardrails, served with NIM |
| Voice and multimodal | Speech in the lab and reasoning over spectra Planned |
| Audit | Every query, retrieval and action logged with user and time |
04 · Place in the loop
Mentra is the front door. It frames the question at the start of the loop and explains each result as it comes back, so a program lead can follow a campaign without reading every table.
Reads the literature and your notebooks, frames the question and drafts the campaign.
Turns the target into candidates and picks the next batch by expected information.
Predicts energies, forces and stresses at near-DFT accuracy, with an uncertainty for each.
Relaxes, scores against the hull and runs GPU molecular dynamics on the survivors.
Sends the short list to partner or in-house labs and brings measurements back.
05 · Questions
It is built to answer only from retrieved sources and to say when it has none. It can still misread a source, which is why every claim carries a link you can check.
Open-access literature to start; licensed publisher and patent corpora depend on agreements we have not signed yet. Your own documents are always available to your tenant.
Not in the first release. We may add per-team budgets for small screens later, if pilot partners ask for it.
Next step
The 2027 pilot pairs one of your programs with our scientists for twelve weeks. You bring a target and reference data; we bring the GPUs and the models.