Split, Croatia — available for new work
Luna AI Lab is an AI research studio that also helps companies with software and process automation. We design and train generative and probabilistic models, then take them the rest of the way: distributed training, evaluation you can defend, and a system that actually runs.
Three things I do well enough to charge for. Everything else I will tell you to hire elsewhere.
Automatizacija poslovnih procesa za tvrtke u Hrvatskoj →
Diffusion, flow matching and functional generative networks for high-dimensional conditional sampling. When iterative denoising is too slow or too expensive to run, I collapse it into single-shot generation without losing ensemble calibration, fidelity or diversity.
Fine-tuning at 70B scale with PEFT/LoRA and 8-bit quantisation, retrieval that returns the right context instead of merely nearby text, and multi-agent setups that hold together under self-play. Typed outputs, and an evaluation harness before anything reaches a user.
A model that only runs in a notebook is not finished. Distributed multi-GPU training, containerised inference, serverless GPU deployment, monitoring and retraining. I have taken a paid product through this entire stack on my own, billing included.
Problem, what got built, and the number it moved.
Co-led the technical design and training of the diffusion and flow-matching models behind GEM-1, then GEM-2 (equal contribution): a ~275M-parameter one-shot extension of diffusion models built on functional generative networks. It jointly models global atmospheric dynamics and the variables people actually decide on — daily maximum and minimum temperature, precipitation, wind extremes — in a single forward pass. I owned the pipeline end to end: architecture, training data, distributed multi-GPU training, evaluation systems and deployment. Co-author on GEM-3, now in production.
Outperforms NOAA GEFS and ECMWF IFS ENS on CRPS skill at 1–40 day leads, with state-of-the-art relative economic value and positive skill maintained to 126 days. Two US AI patents.
A paid audio-transcription service, built alone. WhisperX plus NeMo speaker diarisation packaged as a multi-stage Docker image on RunPod serverless GPUs; Next.js on Vercel; Firebase for auth and Firestore; Cloud Storage for audio and transcripts; Stripe Checkout for credits; Cloud Functions running the yt-dlp and ffmpeg transcode pipeline.
Research, infrastructure, frontend, payments and support — one person, one stack, in production.
A multi-task encoder-decoder transformer over cycle history, symptoms, mood, heart rate, skin temperature and respiratory rate, unified into one pipeline. Shipped to users end to end, from data engineering and GCP monitoring through to how the predictions were presented in the app.
Period start MAE 2.3 days, period end MAE 0.68 days — roughly 50% better than the previous baseline — and ovulation F1 0.922.
Unfamiliar domain, fixed deadline, top-percentile result — repeatedly, and across modalities that share almost nothing. Lumbar spine MRI classification 15th of 1,874. EEG seizure detection 54th of 2,767. Accelerometer sleep-state time series 42nd of 1,877. Vesuvius 3D segmentation 39th of 1,249.
The useful signal is not any single placing. It is that medical imaging, EEG, time series and 3D vision all came out in the same band.
CatBoost and XGBoost regression on encrypted, era-structured financial features with multi-target training, era-aware hyperparameter sweeps and out-of-fold evaluation. Daily inference runs unattended on Cloud Run behind an HTTP trigger.
Proof that not every problem needs a neural network — and that the deployment story matters more than the model choice.
Small, fixed first step. No retainer before there is evidence.
Fixed scope, fixed price, one to two weeks. You bring the problem and whatever historical data exists. You get back a backtested model, an honest evaluation, and a written answer to whether this is worth building at all. Sometimes that answer is no, and this is the cheap way to find out.
Architecture, training, evaluation and deployment of the real system. We agree the metric before I start, and you see progress against it every week rather than at the end.
Once it runs: retraining, drift monitoring, and the changes that follow from what production teaches you. Monthly, and cancellable.
Not in scope 24/7 on-call, on-premise or SCADA integration work, and staff augmentation. If that is what the project actually needs, I will say so on the first call rather than three months in.
Describe the problem in a few sentences. I reply to everything that is not a template.