# Arno Candel

> Building at the Frontier of AI and Physics

- Location: Los Gatos, CA
- Site: https://candel.org
- GitHub: https://github.com/arnocandel

## Profile

Scientific-computing and ML-systems engineer with hands-on experience spanning
simulation-generated data, distributed ML, CUDA kernels, production inference,
and frontier-lab agent evaluation. Former H2O.ai CTO; Member of Technical Staff
at xAI; prior SLAC/ACE3P supercomputing and Skytree (first technical hire).

## Social

- LinkedIn: https://linkedin.com/in/candel
- X: https://x.com/arnocandel
- GitHub: https://github.com/arnocandel
- YouTube: https://www.youtube.com/results?search_query=%22Arno+Candel%22
- ORCID: https://orcid.org/0009-0001-0262-3354
- Kaggle: https://www.kaggle.com/arnocandel

## Experience

### ETH Zurich (Foundations · Physics)

> Masters summa cum laude · PhD on X-ray free-electron laser source simulation — Maxwell's equations, relativistic particle dynamics, HPC

- Physics
  - Masters, summa cum laude
  - PhD (Simulation of electron source for next-generation X-ray free-electron laser)

### SLAC (2005–2011 · Staff Scientist · Supercomputing)

> #1 on ACE3P (~4.7k) — C++/MPI Pic3P particle-in-cell, T2P wakefields, Omega3P/S3P on Jaguar XT5 (#1 on top500.org)

- #1 committer (~4.7k commits) on ACE3P
  - Pic3P/TFE3P — 3D particle-in-cell for RF guns and cavities (LCLS/XL5: beam loading, drive fields, phasespace tools)
  - T2P — time-domain wakefields (monopole/dipole), validated vs ABCI/MAFIA
  - Omega3P/S3P — frequency-domain eigenmode and S-parameter solvers (impedance, ABC, waveguide BCs)
  - Multipacting, mesh conversion, and acdtool postprocessing

### Skytree (2012–2014 · Senior MTS · First technical hire)

> $20M Series A (2013) · acquired by Infosys · C++/MPI/OpenMP — RF/GBM, recommenders, approx NN, streaming k-means

- #1 committer (~3k commits)
  - Large-scale C++/MPI/OpenMP ML server — core table/tree runtime, HDFS/Hadoop I/O, streaming queries, and model I/O
  - Distributed RF/GBM ensembles and SVM
  - Batch and streaming k-means, KDE/KDA
  - Exact and approximate nearest neighbors, metric learning, weighted NN classification
  - Distributed collaborative filtering/recommenders, QA, and cluster benchmarks

### Fortune · Wired · Stage (Spotlight · 2014–)

> Fortune Big Data All-Star · Wired · KDD, GTC, Strata, H2O World

- Named a Fortune Big Data All-Star (2014)
- Featured in ETH Alumni (2015), ETH Globe (2015), and Wired (2017)
- Frequent speaker at KDD, NVIDIA GTC, Strata, H2O World, and Deep Learning conferences

### H2O.ai (2014–2025 · CTO · 12 years)

> #1 on H2O-3, Driverless AI, h2oGPTe — distributed ML → AutoML → GenAI

- Top contributor for the 3 main revenue-generating products at H2O.ai
  - **H2O-3**: #1 committer (~4k commits) — core distributed algorithms (GBM, deep learning, DRF/XGBoost), metrics, UI, Java runtime and scoring, and Python/R APIs for 100k+ community users.
    - https://github.com/h2oai/h2o-3/graphs/contributors?all=1
  - **Driverless AI** (customer testimonial): #1 committer (~16k commits) — led the genetic AutoML engine (feature transformers, model search, ensembles), GPU training (XGBoost/LightGBM/Torch), time-series and NLP/BERT pipelines, and Java production scoring used in banking and telco.
    - https://h2o.ai/platform/ai-cloud/make/h2o-driverless-ai/
  - **h2oGPTe**: #1 committer (~8k commits) — core platform, RAG retrieval, document parse/crawl, agent tooling and file workflows, chat UI, evals/benchmarks, and private/on-prem deploy (Docker/Helm); #1 on GAIA (Dec 2024, Mar 2025).
    - https://h2o.ai/platform/enterprise-h2ogpte/
  - Also #2 contributor to open-source h2oGPT (~780 commits) — Llama/CodeLlama fine-tuning (LoRA/PEFT, bf16, multi-GPU), training-data pipelines, TGI/transformers inference, prompt templates, and GPU performance benchmarks
  - Top-three contributor to open-source H2O4GPU (K-means CUDA) — CUDA/C++ GPU solvers for K-means and GLM (cuBLAS, CUB, Thrust, multi-GPU)
  - Inventor on patent

### xAI (2025 · Member of Technical Staff)

> Macrohard CUA — data loops & Ray-on-Spark video pipelines for pre-training

- Macrohard computer-using agent (CUA) at xAI
  - Data and feedback-loop infrastructure across perception, action selection, post-training, and benchmarks
  - Ray-on-Spark video pipelines for ~90k hours of CUA labeling and captioning for pre-training

## Selected publications

- (2023) H2O Open Ecosystem for State-of-the-art Large Language Models — EMNLP 2023 System Demonstrations
  https://aclanthology.org/2023.emnlp-demo.6/
- (2023) h2oGPT: Democratizing Large Language Models — arXiv
  https://arxiv.org/abs/2306.08161
- (2017) Benchmarks and Process Management in Data Science: Will We Ever Get Over the Mess? — KDD 2017
  https://doi.org/10.1145/3097983.3120998
- (2012) Modeling and design of an X-band rf photoinjector — Physical Review Special Topics – Accelerators and Beams
  https://doi.org/10.1103/PhysRevSTAB.15.102001
- (2011) 50 MW X-Band RF System for a Photoinjector Test Station at LLNL — PAC / conference proceedings
  https://accelconf.web.cern.ch/PAC2011/papers/TUP132.PDF
- (2011) A Reduced Gradient Output Design for SLAC's XL4 X-Band Klystron — PAC / conference proceedings
  https://inspirehep.net/files/131fd5ef7db93e3a078f9ccd93e7cdb4
- (2011) Advances in Parallel Electromagnetic Codes for Accelerator Science and Development
  https://inspirehep.net/files/e2b1d160fe5c7a2c5fb59187d9847109
- (2011) An Optimized X-band Photoinjector Design for the LLNL MEGa-Ray Project — PAC / conference proceedings
  https://accelconf.web.cern.ch/PAC2011/papers/MOP128.PDF
- (2011) LLNL's Precision Compton Scattering Light Source
  https://inspirehep.net/files/fae3bfca9ba429b00c7aba5aa6021e27
- (2011) Numerical Validation of the CLIC/SwissFEL/FERMI Multi Purpose X Band Structure
  https://inspirehep.net/files/c1c59b243e627e203c407a3ee17775b1
- (2011) Numerical Verification of the Power Transfer and Wakefield Coupling in the CLIC Two-Beam Accelerator — PAC / conference proceedings
  https://inspirehep.net/files/ad25b99071807bd14b780c4f8cb10238
- (2011) Precision X-band Linac Technologies for Nuclear Photonics Gamma-ray Sources — PAC / conference proceedings
  https://inspirehep.net/files/642e4a68b7f8251b515ffcdf13b5f621
- (2011) X-Band RF Photoinjector Research and Development at LLNL — PAC / conference proceedings
  https://accelconf.web.cern.ch/PAC2011/papers/TUP023.PDF
- (2011) X-Band Test Station at Lawrence Livermore National Laboratory — PAC / conference proceedings
  https://inspirehep.net/files/ef7f145b3d81b7e769f731201d68c61a
- (2010) ACE3P Computations of Wakefield Coupling in the CLIC Two-Beam Accelerator
  https://inspirehep.net/files/6765c80c9266083df500dfafaa8b40be
- (2010) Dark Current Simulation for  the CLIC T18 High Gradient Structure
  https://inspirehep.net/files/334b6021d91c13d685a37fb9b88dee05
- (2010) On using moving windows in finite element time domain simulation for long accelerator structures — Journal of Computational Physics
  https://doi.org/10.1016/j.jcp.2010.08.037
- (2009) High-Fidelity RF Gun Simulations with the Parallel 3D Finite Element Particle-In-Cell Code Pic3P — AIP Conference Proceedings
  https://doi.org/10.1063/1.3215604
- (2009) Parallel 3D Finite Element Particle-in-Cell Simulations with Pic3P
  https://inspirehep.net/files/4cb1456e00bbbd87096e7be61169291b
- (2009) Parallel Higher-order Finite Element Method for Accurate Field Computations in Wakefield and PIC Simulations
  https://inspirehep.net/files/f09727eaa1f6f94b23fda5899eb45087
- (2009) State of the art in electromagnetic modeling for the Compact Linear Collider — Journal of Physics: Conference Series
  https://doi.org/10.1088/1742-6596/180/1/012004
- (2009) Wakefield Computations for the CLIC PETS using the Parallel Finite Element Time-Domain Code T3P
  https://inspirehep.net/files/03befa4f0058c175902b7271aecf8e4c
- (2009) Wakefield Simulation of CLIC PETS Structure Using Parallel 3D Finite Element Time-Domain Solver T3P
  https://inspirehep.net/files/ad8ec3f81d77213bef9dffa6e09cd38e
- (2008) Computational Science Research in Support of Petascale Electromagnetic Modeling — Journal of Physics: Conference Series
  https://doi.org/10.1088/1742-6596/125/1/012077
- (2008) Design and Optimization of Large Accelerator Systems through High-Fidelity Electromagnetic Simulations — Journal of Physics: Conference Series
  https://doi.org/10.1088/1742-6596/125/1/012003
- (2008) On Projecting Discretized Electromagnetic Fields with Unstructured Grids
  https://inspirehep.net/files/b952ba30c54d432ddf9c194bdef46ef8
- (2008) Parallel 3D Finite Element Numerical Modelling of DC Electron Guns
  https://inspirehep.net/files/9525260ee5dd799969c4d4327821f5a1
- (2008) Parallel Computation of Integrated Electromagnetic, Thermal and Structural Effects for Accelerator Cavities — PAC / conference proceedings
  https://inspirehep.net/files/791534d3a5225a52033405fa579346c7
- (2007) Parallel Finite Element Particle In Cell Code for Simulations of Space charge Dominated Beam Cavity Interactions — PAC / conference proceedings
  https://doi.org/10.1109/PAC.2007.4440757
- (2007) Towards Simulation of Electromagnetics and Beam Physics at the Petascale — PAC / conference proceedings
  https://doi.org/10.1109/PAC.2007.4441117
- (2006) A Massively parallel particle-in-cell code for the simulation of field-emitter based electron sources — Nuclear Instruments and Methods A
  https://doi.org/10.1016/j.nima.2005.11.059
- (2006) State of the Art in EM Field Computation — PAC / conference proceedings
  https://inspirehep.net/files/7939befa6264535672e0ee27bf25b4c2
- (2005) Low Emittance X-FEL Development — eConf
  https://inspirehep.net/files/9594dbc414ad9bc7225effd31938b773
- (2005) Simulation of Electron Source for Next-Generation X-ray Free-Electron Laser — PhD thesis, ETH Zürich
  /papers/phd-thesis.pdf
- (2004) Electron Beam Dynamics Simulations for the Low Emittance Gun
  https://accelconf.web.cern.ch/e04/PAPERS/THPLT018.PDF
- (2004) Low emittance gun project based on field emission
  https://accelconf.web.cern.ch/f04/papers/THPOS01/THPOS01.PDF
- (2004) Preliminary Results on a Low Emittance Gun Based on Field Emission
  https://accelconf.web.cern.ch/e04/PAPERS/MOPKF005.PDF

## Projects (hosted at https://candel.org)

- **Standard Model of Physics** [Physics, Lean]: Arithmetic canonicity of a Lerch encoding for the finite SM+ν_R matter architecture — affine-orbit minimality, Hecke eigenfunctions, finite Weyl algebras, and audited matter interfaces.
- **MadMax AutoML** [AutoML, Tabular] (GPU): Next-gen AutoML for tabular supervised learning — binary/multiclass classification and regression with ensembles, ONNX serving, and a training UI.
- **Exact & Scalable Kelly** [Streamlit, Quant finance]: Portfolio optimization with exact Kelly (sample log-growth) vs mean-variance Kelly, rolling backtests, and benchmarks on real market data.
- **Analytic Ray Tracing** [Streamlit, Telecom] (GPU): Telecom digital-twin demo — analytic ray tracing over a synthetic city geometry with Numba/CUDA acceleration.
- **Analytic Dosimetry** [Streamlit, Dosimetry] (GPU): Spectral RBE coprocessor — translate macroscopic physical dose into biological damage estimates with interactive spectral analysis.
- **Exact Structured Antibiotic Twin** [Streamlit, Biology]: Exact hidden-state twin for antibiotic strategy — compare regimens on viable burden, filamented fraction, and biomass spectrum with exact-vs-grid timing.
- **Analytic Cone-Beam CT Scatter** [Streamlit, CBCT] (GPU): CBCT scatter coprocessor demo — analytic multi-scatter resolvent vs Monte Carlo, SPR maps, and reconstruction artifact correction on a stylized thorax slice.
- **Analytic Dark-Field CT** [Streamlit, CT]: Dark-field imaging coprocessor demo — analytic vs Monte Carlo dark-field signal for lung and breast X-ray CT with interactive ray diagnostics.
- **Cinema Acoustics** [Streamlit, Acoustics] (GPU): Acoura parametric FEM for cinema acoustics — RP22 speaker layout, symmetric subwoofer optimization, and CuPy-accelerated Helmholtz solvers on theater meshes.
- **Casimir Transistor** [Streamlit, Physics]: Interactive Casimir transistor design — finite-T Lifshitz maps, exact planar gate-slab residual, and nonadditivity where pairwise intuition breaks down.
- **Hybrid Tumor Lerch Twin** [Streamlit, Oncology]: Hybrid tumor digital twin with an exact multi-spine Lerch inner solver — clone/size-structured biomarkers, POD benchmarks, and posterior predictive therapy bands.

## Machine-readable discovery

- https://candel.org/llms.txt
- https://candel.org/llms-full.txt
- https://candel.org/resume.md
- https://candel.org/sitemap.xml
- https://candel.org/robots.txt

---
Generated from `src/config.ts` (same `eras` + `credentials` as Explore/Resume).