Viktor Cikojević, PhD

Senior ML research engineer with end-to-end ownership across generative modelling, LLM systems and production deployment. At Salient Predictions I designed and trained the diffusion and flow-matching architectures behind GEM-1, then co-led GEM-2 (arXiv:2601.03753): a ~275M-parameter one-shot extension of diffusion models built on functional generative networks (FGN), cutting training cost ~100× and inference cost ~10× against multi-step diffusion ensembles such as GenCast. Alongside that I have shipped a containerised audio-transcription SaaS on GCP and RunPod, and have hands-on LLM fine-tuning, RAG and multi-agent experience from Kaggle competitions and side projects. Published researcher — 2 US AI patents, 10 Q1 physics papers, ~300 citations — and Kaggle Expert, top 551 of 205,000. PhD in computational physics, with deep PyTorch and distributed multi-GPU training experience.

1. Machine Learning Research Engineer

   Dec 2023 – Feb 2026

   Salient Predictions · Split, Croatia (remote)

   • Co-led the technical design and trained the diffusion and flow-matching generative models powering GEM-1: high-dimensional conditional sampling over global atmospheric state, with controllable, calibrated ensembles.
   • Co-led the technical design and training of GEM-2 (equal contribution, arXiv:2601.03753): a ~275M-parameter one-shot extension of diffusion models via functional generative networks (FGN), jointly modelling global atmospheric dynamics and decision-centric variables (daily Tmax/Tmin, precipitation, wind extremes); outperforms NOAA GEFS and IFS ENS 46 on CRPS skill at 1–40 day leads; state-of-the-art relative economic value (REV); positive skill maintained to 126 days (S2S range).
   • Co-author on GEM-3, Salient's latest production generative model (paper forthcoming).
   • Cut training cost ~100× and inference cost ~10× versus multi-step diffusion ensembles (e.g. GenCast) by collapsing iterative denoising into single-shot generation, while preserving ensemble calibration, fidelity and diversity.
   • Owned the full research-to-production pipeline: architecture design, training-data preparation, distributed multi-GPU training, evaluation systems (CRPS, REV, calibration, loss-pattern diagnostics), and deployment.
   • Co-invented 2 US AI patents, including GEM-2.
   • Delivered the prior-generation production model: +20% skill score over the previous baseline.

2. Data Scientist

   Jan 2022 – Dec 2023

   Bellabeat (Y Combinator backed) · Split, Croatia (remote)

   • Developed a multi-task transformer (encoder-decoder) for menstrual health prediction: period start MAE 2.3 days, period end MAE 0.68 days (~50% improvement over baseline), ovulation F1 0.922; published article.
   • Ingested multimodal inputs (cycle history, symptoms, mood, heart rate, skin temperature, respiratory rate) into a unified ML pipeline.
   • Shipped user-facing health features end to end, from data engineering and production pipelines, monitoring and governance on GCP through to report design, working with product and design to put model outputs in front of users.

3. Computer Vision Engineer

   Jul 2021 – Jan 2022

   Codeasy / Necogi · Split, Croatia

   • Built an aerial imagery analysis pipeline in PyTorch for object detection and classification.

4. PhD in Computational Quantum Physics (with honours)

   Dec 2016 – Oct 2021

   UPC Barcelona & University of Zagreb

   • Developed high-performance Quantum Monte Carlo simulation software in C++ for probabilistic sampling.
   • Built Python pipelines for large-scale data analysis; published 10 Q1 papers with ~300 citations.
   • PhD thesis · Google Scholar

5. MSc in Computational Physics — CGPA 5.0/5.0

   Sep 2014 – Sep 2016

   University of Split · Split, Croatia

   • Coursework in computational simulations; master's thesis on quantum Monte Carlo, with an Erasmus exchange at UPC Barcelona.

Machine learning. P. Rauba*, V. Cikojević*, F. Bartolić*, S. Levang* et al., “Probabilistic Transformers for Joint Modeling of Global Weather Dynamics and Decision-Centric Variables,” arXiv:2601.03753, 2026 (*equal contribution). Read on arXiv. GEM-3 paper forthcoming.

Physics. 10 peer-reviewed publications in Physical Review A/B/Research and New Journal of Physics, ~300 citations. Google Scholar.

• RSNA 2024 — Lumbar Spine Degenerative Classification. Medical imaging (MRI). 15th of 1,874, top 1% — solution writeup.
• HMS — Harmful Brain Activity Classification. EEG-based seizure detection. 54th of 2,767, top 2%.
• Detect Sleep States. Time-series accelerometer data. 42nd of 1,877, top 2%.
• Further results. Vesuvius Challenge (computer vision / 3D segmentation, 39th of 1,249, top 3%), LLM 20 Questions (LLM agents, 87th of 832), ISIC 2024 (computer vision / medical imaging, 243rd of 2,739).

• transcribevoice.app & transcribe-api-endpoint — production audio-transcription SaaS. Containerised WhisperX + NeMo (speaker diarisation) inference service with a multi-stage Docker image, deployed on RunPod serverless. Next.js frontend on Vercel, Firebase Auth/Firestore, Google Cloud Storage for audio and transcripts, Stripe Checkout for credits, GCP Cloud Functions for the upload and transcoding pipeline (yt-dlp + ffmpeg).
• Kaggle LLM Science Exam — closed-domain QA over Wikipedia. Retrieval pipeline with sentence-transformers + FAISS for context selection; fine-tuned DeBERTa and 70B-scale LLaMA/Platypus with PEFT/LoRA and 8-bit quantisation (bitsandbytes); Pydantic-typed LLM outputs for synthetic data generation.
• Kaggle LLM 20 Questions — multi-agent self-play. Questioner/answerer agent pair using LangChain with a Chroma vector store, and DSPy; ColBERT for retrieval; local inference via llama-cpp-python and Ollama; self-play orchestrated across category templates.
• Numerai — classical ML in production. CatBoost/XGBoost regression on encrypted, era-structured financial features with multi-target training; era-aware hyperparameter sweeps and OOF evaluation; daily inference deployed on GCP Cloud Run behind an HTTP trigger.

Languages

Python, C/C++, SQL, Bash

Generative modelling

Diffusion models, flow matching, functional generative networks (FGN), one-shot generation, probabilistic transformers, conditional sampling, ensemble calibration

ML / deep learning

PyTorch, distributed multi-GPU training, transformers, TensorFlow, Keras, CNNs, LSTMs, computer vision, NLP

Classical ML

XGBoost, CatBoost, scikit-learn, tree ensembles, feature engineering, hyperparameter optimisation, model evaluation

LLM & agents

LLM fine-tuning (PEFT/LoRA, 8-bit quantisation with bitsandbytes), RAG (sentence-transformers, FAISS, Chroma, ColBERT), LangChain, DSPy, Chainlit, multi-agent self-play, Pydantic-typed outputs, llama-cpp-python, Ollama

Production / MLOps

Docker (multi-stage builds), RunPod serverless, GCP (Cloud Run, Cloud Functions, Cloud Storage, BigQuery), Firebase (Auth, Firestore), Vercel/Next.js integration, Git/GitHub, GitLab

Data tooling

xarray, dask, Pandas, NumPy, WandB, Hydra

Running and swimming, guitar, travelling.

For consulting and project work, see Luna AI Lab or email viktor.cikojevic@luna-ai-lab.com.