Vergence / AI-native optical design

Meet your new copilot

An optical bench.
A copilot that
thinks with you.

Explore the copilot Join the beta

Somewhere to work a lens design out.

Ask for a change in plain language and read the diff before it applies. The copilot does the reasoning; the ray-trace engine produces every number it quotes.

Inspect the reference lens
Download .zmx
every number below came out of the engineStart with the light

A live ray trace through the reference lens

meridional trace · axial and field bundles · F, d and C 8.0°
50.022 mm
effective focal length
F/5.00
working f-number
4.97 µm
rms spot · on axis
9.03 µm
rms spot · at field

Recorded in the Vergence engine. Explore the field, inspect the light. View the source data ↗

Meet VergenceAI

Ask it what you would ask
the person at the next bench.

It has the same tools you do. Every answer names the ones it ran and shows you what they returned.

Vergence / VergenceAIInteractive feature walkthrough
Find precedentStart with what is knownBuild a planMake the next move explicitReview the evidenceConnect findings to physicsUnderstand the whyDevelop your optical intuition
Your call

Read the proposal.
Approve the next step.

Retrieval-grounded reasoning

Reference / Cooke triplet

Start from a lens that already works.

Find designs with a comparable spec and structure, then open one and see how it was solved.

You

Where should I start with a compact visible-light objective?

VergenceAI

Explore a Cooke triplet as a reference. Inspect its first-order properties, glass choices, and field performance before deciding whether it fits your spec.

Reference designCooke tripletInspect ↗
RefractiveVisible spectrumCatalog glass
Effective focal length50.022 mm
Open the reference prescription ↓

The precedent is there to argue with, not to copy.

Natural language → staged optimization

Proposal / your decision

A plan you can read before it runs.

What you asked for becomes stages, each with its own targets, variables, constraints and budget. Edit any of it before anything runs.

You

Help me refine this objective. Keep the focal length controlled as we improve image quality.

Establish the first-order targets

Begin with a constrained set of variables. Set the focal-length target and physical bounds before widening the search.

Open the design space

Unlock additional radii and spacings. Inspect the merit function, search budget, and constraints for this stage.

Refine and inspect the result

Review the stage metrics and saved checkpoints. Decide which result to adopt into your working design.

The plan is a proposal.Nothing runs until you approve it.

Evidence-backed design review

Recorded engine output

Every finding names its evidence.

The review is written from an evidence pack the engine computed first. Open any finding to see the analysis underneath it.

Supporting evidence

50.022 mm

Effective focal length

A first-order property of the reference lens. Changing the viewing field does not change its focal length.

On axis4.97 µm
At 8°9.03 µm

Recorded RMS spot sizes from the same lens at different fields. Use the ray-trace slider above to explore the field dependence.

Inspect the sourcepython backend/tools/build_landing_data.py → landing/public/cooke.jsonView recorded engine data ↗

The report is tied to the exact design it was written about, and keeps its evidence with it.

Socratic tutor

Observe → hypothesize → test

Ask it to teach instead of answer.

Same tools, different reply. You get a question back, a hint if you want one, and an analysis to go and run.

You

Why does the spot change when I move off axis?

Tutor

Start with an observation: move the field slider on the reference lens. What changes in the spot-size readout, and what stays the same?

Explore the reference lens ↑
Give me a hint

The focal length belongs to the system. Image quality can vary with field. Compare the axial and field bundles, then inspect a spot diagram or ray fan to investigate the aberrations.

What should I investigate next?

Ask whether the blur is symmetric or directional, and how it changes across wavelength. Those observations help distinguish possible causes; a spot-size value alone does not identify the aberration.

It guides while you are working something out, and answers straight when you ask it a fact.

Illustrative conversations · recorded reference measurementsExplore the evidence ↗

It cites precedent.Related designs, named, with their sources.

The engine measures.Every figure in an answer came out of an analysis.

You approve.A change arrives as a diff you read first.

The design loop

Follow the light.
The evidence follows it.

  1. S1 describe

    Say it, or hand it a patent.

    Ask in the language you would use with a colleague, or drop in a patent PDF and let the extractor read the prescription tables off it. Scans work too.

    a fast triplet for the visible, distortion under one percent
    propose_merit_function

  2. S2 trace

    The engine builds it and traces it.

    Catalog glass on the dispersion fit its vendor published, the apertures the design actually has, and the conjugates you specified. Nothing is approximated to make a picture come out.

    F2 · resolved to schott, not cdgm
    nd 1.62004 · Vd 36.37 · in band

  3. S3 verify

    Nothing reaches you unchecked.

    A deterministic verifier rebuilds what was extracted and compares it against what the document claims about itself. Anything that survives that is citable, and anything that does not comes back refused, with the reason on screen.

    stated focal length · rebuilt and compared
    stated f-number · rebuilt and compared
    disposition ACCEPT · nothing left unchecked

AI for the thinking.
Physics for the proof.

VergenceAI reads your design and tells you what it makes of it. The numbers in that answer come from the engine, and every one of them traces back to an analysis you can open, rerun and export yourself.

It proposes; you decide what gets applied.

Verified import

We tell you when we cannot reproduce a lens.

Hand Vergence a patent and it reads the prescription, rebuilds the system and checks it against what the document claims about itself. This is how that came out across 82 embodiments, refusals included.

20
accept

Rebuilt, traced, and agreeing with everything the document states about itself.

17
flagged

Built, but something disagrees or nothing was left to check it against. A person decides, with the reason on screen.

45
reject

The claim could not be reproduced. You are told, and nothing is handed to you as a design.

Why the refusals were refused

  • the rebuilt lens disagreed with a quantity the document states 26
  • the prescription as printed would not trace 17
  • no embodiment could be read from the document at all 6

Bad lenses accepted: zero. Every disposition above is machine-readable and carries its own reason, produced by tools/verify_run.py.

Evidence

Check any of it yourself.

Every figure here exists in the repository and came out of a command that can be run again. None of it is an estimate, and each one is printed with the command that produced it.

What was measured, and what produced it Result
First-order agreement with OpticStudio, 33 clean refractive designs tools/opticstudio_check.py
0.0074 %
Tolerance sensitivity against the closed form df/dR₁ app/tolerancing/sensitivity.py
0.0351 %
Monte Carlo σ against the closed form app/tolerancing/montecarlo.py
0.3584 %
Every analysis over every corpus design, 105 of them tools/parity_sweep.py
1785 runs · 0 fail · 0 crash
Zemax .zmx files imported, none failing tools/import_sweep.py
100 / 100 · 60 clean · 40 reported
Export round trip across the corpus tools/roundtrip_sweep.py
95 exact · 10 reported · 0 silent drift

Not yet proven: an independent Zemax column measured by a third party, design by design. It is in progress, and we will publish it whole, including the rows where we lose.

Platform

The whole design loop, in a browser tab.

There is no licence server to check out and no workstation to sit at. It runs on whatever machine you already have.

Design and optimize
Local and global search over a merit function you write and inspect, with real catalog glass and discrete glass substitution.
25 merit-function operands
Analyze
Spot, wavefront, MTF, distortion, field curvature, colour, illumination, encircled energy, and Seidel attribution per surface.
20 analyses in the registry
Tolerance
Deterministic sensitivity that names the surface and the parameter costing you most, then Monte Carlo yield over thousands of perturbed builds.
6 tolerance types
Report
A prescription report computed on the server, where every value is assembled from an engine call and nothing is calculated in the document itself.
Interop
Import and export Zemax .zmx, with a translation report that names anything it could not carry across rather than quietly dropping it.
1989 glasses in the catalog
VergenceAI
Reads your system, proposes changes behind an approval gate, and applies nothing you have not seen as a diff first.
34 VergenceAI tools

Who it is for

Built for people who will check.

Thermal and IR

Germanium, silicon, zinc selenide and the chalcogenides resolve to the fit their vendor published. Outside the band that fit covers, Vergence refuses the glass by name instead of extrapolating it.

Machine vision

Compact imaging lenses from a spec to a toleranced prescription, with a yield number you can hand to a supplier.

Students

The same tools a working designer uses, with a tutor mode that answers a question with a question and points you at the analysis that settles it.

Research labs

Fast iteration on one-off instruments, and reports whose numbers you can put in a paper and defend.

The image plane

Bring your next idea
into focus.

The alpha is invite-only and deliberately small for now. Leave an address and we will write to you when there is room.

one message when a seat opens, and nothing else ever

We store your address to send one invitation and nothing else. Our privacy notice.