Skip to main content

Run what exists Adapt what works Return the evidence

Primae runs open-source scientific software before choosing a starting point, then adapts the best-evidenced foundation to the problem.

Request a fit review Review recorded evidence
Run candidates before choosing
03 / 03
Code and evidence travel together

The repository returns with passed, failed, and unresolved checks attached for qualified review.

Primae runs shortlisted open-source candidates before choosing a starting codebase, adapts the selected foundation, executes problem-specific checks, and returns working code with a signed record that keeps failures and unresolved work visible for human review.

Not an AI scientist
Not a blank-sheet code generator

Primae works on the scientific software itself

It does not center idea generation, manuscript production, or a plausible patch. It begins with prior implementations, observes them running, changes what the case requires, and evaluates the result against evidence outside the model.

01INPUT

Defined technical problem

Environment, constraints, expected result, and acceptance basis.

02UNIT OF WORK

Executable codebase

A running foundation selected and adapted through execution.

03RETURN

Code + check record

Measured outcomes, limitations, and unresolved items attached.

BEST FIT

Bring a problem that reality can judge

  1. 01
    A defined problem

    Objective, constraints, environment, and expected output.

  2. 02
    An executable result

    Code, a model, simulation, or another testable artifact.

  3. 03
    An external basis

    Reference, equation, invariant, benchmark, dataset, or standard.

A shortlist becomes useful only after it runs

Primae keeps one technical brief connected while the starting point moves from a promising description to observed execution, targeted adaptation, and a reviewable return.

ONE JOB · SIX PUBLIC STAGESProblem · environment · acceptance criteria
06 / RETURN

Code and evidence travel together

Hand back the repository, signed check record, failed checks, unresolved items, and review state.

Primae discovers relevant open-source candidates, runs them, chooses a starting point from observed behavior, adapts the selected foundation, executes problem-derived checks, and returns the repository with a signed record for human review.

Keep the years
Change the delta

An established scientific repository can contain years of domain decisions. Primae preserves the parts that already work and concentrates new work on the gap between the running foundation and the technical case.

Economic objective · Make effort follow the missing change—not a blank-sheet rebuild.

The selected repository remains the foundation. Primae targets the missing capability and keeps its contribution distinguishable from the code that was already there.

Checks are written from the science, then run

An AI opinion is not a check

In staging, Primae supports roughly eighteen families of problem-specific checks. The applicable battery depends on the case, executable environment, and evidence available.

DIMENSIONAL CONSISTENCYCONSERVATIONCONVERGENCEREFERENCE DISTANCE INPUT SENSITIVITYEDGE ROBUSTNESSMODEL AGREEMENTSTANDARDS-LINKED CRITERIA

Failed is a result, too

Each check can carry the measured value, expected value, tolerance, deviation, and confidence. The record remains bound to the checked artifact and preserves outcomes that failed or could not be completed.

A candidate is never silently promoted to an answer.

SIGNED CHECK RECORDOUTCOME STATES
Unresolved stays unresolved

A missing basis or incomplete execution remains visible; it does not become a green result.

A Primae check record can preserve passed, failed, and unresolved outcomes alongside measured and expected values, tolerances, deviations, confidence, the checked code fingerprint, time, and exact check list.

It can work while you sleep
It cannot accept its own work

Primae can continue long-running work, reconnect to the editor, and re-check a shadow copy. A bounded improvement can return with its rationale, but nothing reaches the working repository until a person accepts it.

  1. 01Continue

    The job remains owned and recoverable when the editor disconnects.

  2. 02Show the work

    Discovery, execution, measured checks, and handoff stream into the editor.

  3. 03Try safely

    Bounded attempts run on a shadow copy and can be reverted.

  4. 04Ask for acceptance

    A suggestion and signed rationale return; consequential approval remains human.

SHADOW COPYHUMAN ACCEPTANCEnothing crosses silently

Three recorded problems
No universal benchmark claim

These internal records show how Primae-generated components were compared with named reference implementations across depth estimation, pore-scale CFD, and road-damage detection.

RECORDED COMPARISON / 01

Monocular depth estimation

Nineteen output-affecting components evaluated against the Depth-Anything-V2 reference. Eight DINOv2 backbone modules were recorded separately.

Reference project: Depth-Anything-V2

10matching

1close

1divergent

7not available

RECORDED COMPARISON / 02

Pore-scale CFD

Fourteen physics components reported as coarse categorical verdicts against the MassTransferFoam/OpenFOAM reference.

Reference project: GeoChemFoam / MassTransferFoam

9matching

1close

4divergent

0not available

RECORDED COMPARISON / 03

Road-damage detection

Fourteen output-determining components evaluated against the named yolov5_SODRv1 reference.

Reference project: yolov5_SODRv1

12matching

0close

0divergent

2not available

General machinery
Demonstrated depth in three problem types

The end-to-end staging pipeline has run across machine learning, compiled multiphase simulation, and computer vision. New engagements are assessed case by case; this is not a claim of universal domain or toolchain support.

01

Physical engineering

Mechanics, fluids, structures, materials, controls, and simulation.

02

Life science and chemistry

Molecular modelling, chemistry, biological systems, and scientific software.

03

Earth and energy

Subsurface, climate, geospatial, energy, and environmental models.

04

Scientific ML and numerics

Surrogates, solvers, optimisation, and numerical methods.

NOW

Configured staging engagements

Problem fit, environment, evidence, deployment, and review criteria are established during onboarding.

DATA

Confirm handling first

Deployment and data-handling terms are confirmed before work begins. Do not send confidential source code or datasets through the access form.

REVIEW

Evidence supports the decision

Primae returns code, executed checks, and unresolved work; qualified acceptance remains human.

Bring us a problem with something real to check against

Tell us the problem, current toolchain, and the reference or acceptance criteria you would use. We will review fit and follow up about a design-partner conversation.

  • 01A defined technical problem
  • 02An executable result in mind
  • 03An independent evaluation basis

This is a fit review, not instant enrollment. Pricing, access terms, and service levels are not yet published.

PRIMAE FIT REVIEW3 required · 2 optional

If the form is unavailable, email a non-confidential summary to engineering@singulusai.com. See how form submissions are handled in our Privacy notice.