Hackathon project · 25 October 2026

BEAMTEAMSelf-Driving Optical Engineering with Multiagent Teams

Built with Gemini Argon in the OMP harness and Claude Code, working together with a team of subagents.

Dave Rauchwerk1, Claude Opus 5.5, Gemini 4 Argon

1Independent Researcher

Frame from the BEAMTEAM demo video
Demo video coming soon
Muted loop of the 60-second BEAMTEAM demo. Press Watch with narration for the full version with sound and captions (narration by Gemini TTS). Before the video loads, the poster is a frame taken at 30 s.

TL;DR  Type a topic, pick a figure from an arXiv paper, and BEAMTEAM turns it into a laser-optics bench in under three minutes. Vision and synthesizer agents (Gemini gemini-4-argon) propose the layout. A deterministic Gaussian-beam/Jones tracer verifies every beam. The output is aligned, rendered and exported as a build kit: BOM, printable keyed baseplate and an assembly PDF.

Abstract

Laying out a tabletop laser-optics bench for an atomic or quantum physics experiment is weeks of careful manual work, and the designs that labs publish are usually just schematic figures. BEAMTEAM is a multi-agent system that designs these setups from a natural-language goal, or reproduces them from a figure in a paper. Three vision agents read the figure in parallel. A consensus vote with millimetre scaling turns their readings into a component/beam graph. Two synthesizer agents then build a design in a JSON intermediate representation.

Every design is checked by a deterministic physics tracer. It models Gaussian beams (ABCD/q), Jones polarization, power, AOM orders, fiber mode overlap and clipping. A verifier flags stray beams (laser safety), collisions, unplaced parts and unmet targets. A deterministic autofix and an LLM repair agent close the loop under a hard deadline. Verified designs are aligned by a deterministic tuner, rendered as journal covers and published as a Typst build document and build kit. The kit includes a bill of materials with part numbers, watertight keyed baseplate STLs, and a script for the PyOpticL code-to-CAD library.

Live, paper search to build-document PDF took 1:37 against a 3:00 budget. Three live figure-to-design runs on optical tweezers took 55–74 s and all passed. Grounding the model in real PyOpticL designs cut hallucinated API references from 95 % to 0 %. A 121-optic apparatus traces and verifies in about 6 ms. We also report where it fails: on 8 open-access figures, 4/8 reproductions pass verification but only 3/8 are physically faithful.

How it works

The LLMs propose, the physics disposes.

Architecture diagram: three entry paths (reproduce from a paper figure, goal to design, hierarchical apparatus) feed a shared verify and repair loop, then a tuner, cover artist, publisher and build-kit generator.
Figure 1. Architecture. Gemini LLM agents are blue rounded boxes and deterministic code is grey; stacked boxes are parallel instances. There are three entry paths. Reproduce: paper search, figure fetch, 3 parallel vision agents, consensus merge with mm scaling and 2 parallel synthesizers. Goal to design: a planner and N = 4 parallel designers (best-of-N). Hierarchical: an architect, M parallel module agents and a merge under a shared laser power budget. The reproduce and goal paths share a verify–repair loop: deterministic trace/verify/autofix plus an LLM repair agent (at most 3 rounds). The verified design then goes to a deterministic tuner, an image-model cover artist, a Typst publisher and a build-kit generator.
  1. Find a paperLive arXiv search ranks each paper's figures by how much they look like an optical setup.
  2. Read the figureThree vision agents read it in parallel. A consensus vote and millimetre scaling turn their readings into a component/beam graph.
  3. Synthesize & verifyTwo synthesizers emit the JSON IR. The tracer checks every beam; the verifier flags stray beams, collisions and unmet targets. Autofix adds validated beam dumps and a repair agent fixes the rest, under a hard 115 s deadline.
  4. Fly-in & alignParts drop from the drawing onto the 3D bench and the beams switch on. The bench is misaligned "as built", then a tuner walks every knob back to optimum.
  5. PublishA journal-cover render (Nano Banana 2) and a Typst build-document PDF: plan view, verification, BOM, 3D-print list, assembly, alignment and laser-safety notes.
  6. Build kitBOM.csv, tiled baseplate STLs with keyed alignment-pin holes and beam-path grooves, a PyOpticL/FreeCAD script and an assembly README.

It also supports goal-to-design with best-of-N candidates and hierarchical design of whole apparatus: an architect plus parallel module agents with shared laser power budgets, e.g. a 121-optic Sr+ trapped-ion laser system. Other features are an agent console with colour-coded token streams per agent, and all-optical logic (SOA NAND, χ(2) sum-frequency half adder).

Per-agent timelines of two live reproduce runs on the dual-trap optical tweezers figure.
Figure 2. Two live runs, agent by agent. Reproduce-from-figure on the dual-trap optical-tweezers figure (Gieseler et al., arXiv:2004.05246, Fig. 5), using gemini-4-argon throughout. (a) Three parallel vision passes (23.3–28.1 s) agree 3/3 on 16 parts. The two synthesizers score 0 and 86, and one repair round raises the score to 96. Total time is 89.2 s with 14/14 figure components matched. (b) Deterministic autofix lifts synthesizer 2 from 35 to 80. Both repair rounds regress to 45 and are rejected, so the best design (80, PASS) is kept at 118.4 s. Light segments show time to first token; ticks mark deterministic trace/verify/autofix calls, each under 4 ms.

Key results

Measured on 25 October 2026 on live Gemini calls. Raw runs are in demo/.

1:37
paper search → build-document PDF, live, end to end (budget 3:00)
3/3
live tweezers figure → verified design, 55 / 67 / 74 s, all PASS
95%→0%
PyOpticL API hallucination, zero-shot → grounded
~6 ms
trace + verify of a 121-optic apparatus
4/8 · 4/4
figure benchmark passes verify (3/8 physically faithful) · goal designs pass
Paper search → PDF, live, end to end1:37 (budget 3:00)
Figure → verified design (tweezers, 3 live runs)55 / 67 / 74 s, all PASS
Demo-figure reproduction reliability3/3 PASS, 41–55 s (concurrent)
Goal → verified design~45–85 s, scores 93–98
Paper URL → bench (tweezers, arXiv:2004.05246)PASS 95.4 in 73 s, 14/14 matched
Hierarchical: 6 modules in parallel96 optics, PASS 98.5, ~100 s
Alignment tuner, detuned AOM double-pass102 % of fiber power, 75 steps / 35 ms
Build kit, AOM double-pass18 parts + 58 hardware, ≈ $4.1k (est.), 2 tiles

Reproducing published figures: an 8-figure stress test

Bar chart of wall time and verifier score for reproduce and design modes on 8 open-access figures.
Figure 3. Benchmark. Bars show wall time; labels give the verifier score (green PASS, red FAIL). Reproduce: 4/8 PASS: Raman 96 (46 s), pump–probe 100 (55 s), PDH lock 95 (41 s), and the 22-element dual-trap tweezers 96 (80 s). Raman passes verification, but its physics is wrong: the 532 nm line reaches the CCD. That leaves 3/8 physically faithful. Sr+, MOT/LCVR and HOM fail; dense Yb+ hit the 60 s vision deadline twice. Goal design: 4/4 completed runs PASS. One MOT design run was aborted by a dev-server reload and is not counted.
FigureDiff.ReproduceDesign
Raman backscattereasy96 · 46 s†–
Pump–probe delayeasy100 · 55 s–
PDH thin-disk lockmed.95 · 41 s100 · 52 s
Sr+ ion-trap lasersmed.0 · 104 s75 · 68 s
MOT / LCVR laser systemmed.0 · 90 saborted (infra)
Dual-trap tweezershard96 · 80 s99.5 · 84 s
HOM tomography (3D)hard18 · 119 s–
Yb+ clock laser (dense)harderror ×2–
MALDI (goal only)med.–99 · 50 s

Score (0–100) · wall time; green = PASS. †Passes verify, but the long-pass filter is modelled as a dichroic that lets the 532 nm line through to the CCD. Source: demo/figures/RESULTS.md, which also lists failure patterns and fixes.

Grounding removes API hallucination

Bar chart: hallucinated PyOpticL API references drop from about 95 to 98 percent zero-shot to 0 percent with grounding.
Figure 4. Grounding. Share of PyOpticL API references in gemini-4-argon goal→code output that do not exist. Zero-shot: 94.7 % (AOM, low thinking), 98.1 % (AOM, high), 97.7 % (SAS, low). Grounded with ~17k tokens of real designs plus a signature digest: 0 % in all 4 runs (135/0, 140/0, 130/0, 129/0 valid/invalid refs).
Log-log plot of trace plus verify wall time against number of elements in the design.
Figure 5. Physics is cheap. Deterministic trace + verify time against design size (median of 5; bars min–max). Blue: an 11-element AOM double-pass tiled 1–33× (11–363 elements), all PASS, 0.4–104.6 ms, log–log slope about 1.6. Orange: the 13 example designs; the 121-element Sr+ apparatus took 9.6 ms here under load. On an idle machine it took 6.1 ms. Absolute times carry about ±2× load noise.

Whole apparatus, and light that computes

Plan view of a 121-element Sr+ ion-trap laser apparatus split into 7 module regions, beams coloured by wavelength.
Figure 6. A 121-optic apparatus, designed hierarchically. Plan view of the Sr+ ion-trap laser apparatus, drawn directly from the deterministic trace. It has 121 elements, 11 sources, 134 beam segments and 6 wavelengths (405–1091.8 nm) across 7 module regions on a 1200 × 900 mm table. Beam colour shows wavelength and line width scales with √power. The design passes verification (99.2), with 236.9 of 280.0 mW input delivered.
Truth tables of an all-optical half adder and a sum-frequency AND gate.
Figure 7. 1 + 1 = 10, computed by photons. (a) All-optical half adder on one PPLN crystal: CARRY (631 nm sum-frequency light) is 347.4 mW only for input 11. SUM (residual IR) is correct at a 40 mW threshold. (b) A sum-frequency AND gate outputs 35.3 mW only for input 11. Textbook steady-state models with optimistic conversion efficiency.
Screenshot of the alignment tuning tour: camera at an optic with a live power meter.
Alignment tour. The bench is detuned "as built", then the deterministic tuner walks every knob back. On the AOM double-pass, delivered fiber power goes from 32 % to 102 % of design in 75 trials.

Limitations (honest caveats)

  • Simplified physics. The table model is planar (2D). Gaussian beams are round, lenses are thin, and there is no interference or coherence.
  • Nonlinear elements are textbook steady-state models: SOA gain saturation and χ(2) SFG with perfect phase matching, CW only. The SOA NAND has about 5 dB contrast and is not cascadable in our model. The half adder uses one PPLN crystal with optimistic η.
  • Reproduction works on clean schematics only. Dense multi-panel and 3D figures still fail (Yb+, HOM, MOT). A verifier PASS is not proof of correct physics: the Raman design passes while sending the laser line to the CCD. There is no "sample" or retro-scatter primitive.
  • Notes are not always respected. Focal lengths and laser power given in the notes were sometimes replaced with "convenient" values. Hierarchical designs can come out as disconnected modules with one source each.
  • Generated BOMs are estimates. Prices need checking with vendors, and nothing here replaces a laser-safety review.

Built on PyOpticL & open figures

PyOpticL paper Figure 2a: Rb saturated absorption spectroscopy layout

PyOpticL (Myers et al., Quantum 2026; UMass Ion Trappers) is the code-to-CAD optics concept and library that BEAMTEAM's exporter targets, and our examples are modelled on its designs. It is referenced, not vendored. Left: PyOpticL paper Fig. 2a (Rb saturated absorption spectroscopy, CC BY 4.0), which we used as a test input. Thank you to the PyOpticL authors. BEAMTEAM would not exist without their work.

The benchmark inputs are open-access published figures, credited below as listed in demo/figures/catalog.json. They keep their original licenses. Figures from other arXiv papers, such as Gieseler et al., arXiv:2004.05246, are fetched live at runtime and are not redistributed.

FigureSourceLicense
Raman backscatterToommm (Wikimedia Commons), adapted from T. Schmid & P. Dariz, Heritage 2(2), 1662–1683 (2019) · sourceCC BY 4.0
Pump–probe delayJoachim Terschluesen, "Pump probe setup.svg", Wikimedia Commons · sourceCC0 1.0
PDH thin-disk lockM. Zeyen, L. Affolter, M. A. Ahmed, et al., Rev. Sci. Instrum. (2023) · arXiv:2210.05501CC BY 4.0
Sr+ / Th3+ ion-trap lasersG. Zitzer, J. Tiedau, M. V. Okhapkin, et al., Phys. Rev. A 109, 033116 (2024), Fig. 2 · arXiv:2312.05106CC BY 4.0
MOT / LCVR laser systemB. Piest, V. Vollenkemper, J. Böhm, et al. · arXiv:2109.14277CC BY 4.0
Dual-trap optical tweezersM. A. Ashraf & P. Pullarkat · arXiv:2411.16256CC BY 4.0
Yb+ clock laserY. Zuo, S. Cao, S. Dai, et al. · arXiv:2303.00361CC BY 4.0
HOM tomographyY. Tsujimoto, R. Ikuta, K. Wakui, et al., Phys. Rev. Applied 19 · arXiv:2205.11160CC BY 4.0
Rb SAS (Fig. 2a)PyOpticL paper, Myers et al., Quantum 2026 · PyOpticLCC BY 4.0

What was built at the hackathon vs. what existed

Reproduced from the repository's Provenance section.

Built during the hackathon

2026-10-25, 11:23 → present (see git log)

Everything in this repository:

  • the design IR and API contract;
  • the planar Gaussian-beam/Jones tracer, including nonlinear elements (SOA, SFG, saturable absorber);
  • the verifier, autofix, alignment tuner and module merge;
  • all agent pipelines: goal → design, figure → design, hierarchical, paper search, cover renders, console events and token streaming;
  • the PyOpticL exporter;
  • the BOM, tiled STL baseplates and the Typst build document;
  • the web application;
  • all examples, benchmarks, recorded runs and figures, the paper and the demo video.

The code was written by an AI coding assistant (Claude Code) and its sub-agents, under the team's direction, during the event.

Pre-existing, used but not built by us
  • PyOpticL (Myers et al., Quantum 2026): the code-to-CAD optics concept and library that our exporter targets and our examples are modelled on. Not vendored; referenced only.
  • Google Gemini models via API: gemini-4-argon (agents); Nano Banana 2 / 2.1 (gemini-3.1-flash-image, gemini-nano-banana-2.1) for renders; Gemini TTS (demo narration).
  • Open-source libraries: three.js, React, @react-three/fiber, @react-three/drei, zustand, lucide-react, Vite, TypeScript; FastAPI, Pydantic, NumPy, google-genai, httpx, sse-starlette, uvicorn, manifold3d, trimesh, Pillow; Typst; ffmpeg.
  • Published figures used as test inputs: open-access CC BY / CC0 figures with attribution in demo/figures/catalog.json, plus the PyOpticL paper Fig. 2a (CC BY 4.0). Figures from other arXiv papers (e.g. Gieseler et al., arXiv:2004.05246) are fetched live at runtime and not redistributed.

BibTeX

@misc{beamteam2026,
  title        = {{BEAMTEAM}: Self-Driving Optical Engineering
                  with Multiagent Teams},
  author       = {Rauchwerk, Dave and {Claude Opus 5.5} and {Gemini 4 Argon}},
  year         = {2026},
  howpublished = {\url{https://github.com/elevenarms/beamteam}},
  note         = {Hackathon project}
}

Acknowledgements

We thank the authors of PyOpticL (UMass Ion Trappers) for the code-to-CAD optics library and paper that inspired this project. We also thank the authors of every open-access figure listed above for licensing their work so others can build on it.