Viktor Cikojevic

Salient Predictions built AI weather models. A top research lab acquired the company in around June 2026. I owned the generative modeling line across all three generations of our GEM models, spanning architecture, data preparation, multi-GPU training, evaluation, and deployment. The models forecast global weather as calibrated ensembles, many plausible futures instead of a single guess:

• GEM-1 used diffusion and flow matching.
• GEM-2 (https://arxiv.org/abs/2601.03753), which I co-led with equal contribution on the paper, collapsed the slow multi-step sampling into a single forward pass, while beating operational forecast systems.
• GEM-3 (https://arxiv.org/abs/2608.06241), which I co-authored, trains one set of weights whose forecast timestep is chosen at inference.

I also co-invented two US AI patents, one for GEM-1 and one for GEM-2.

Two Kaggle competitions, one theme: making modern LLM systems run offline on a single 16 GB GPU under strict time limits. In LLM 20 Questions (87th of 832, 2-person team), our agent played every role in the guessing game. A 4-bit Llama-3-8B asks questions by walking binary decision trees we generated offline and audited by hand, then answers by majority voting across three generations grounded in a local ColBERT index of 600,000 Wikipedia abstracts, compressed to 1.6 GB. With no training data, we evaluated by self-play.

In the LLM Science Exam, the questions were written by GPT-3.5, and only 200 labels were given, so we generated around 160,000 synthetic training questions from clustered Wikipedia pages and fine-tuned DeBERTa answer models behind a FAISS retrieval pipeline. To add a much bigger judge, I ran a 70B Llama on one T4 by streaming it layer by layer from disk, reusing the KV cache across the five answer options.

Transcribevoice.app turns audio, video, and social-media links into word-timestamped, speaker-labeled transcripts. I built it solo and ran it in production.

The GPU worker went through eight versions on the RunPod serverless. Speaker labels come from my own diarization pipeline: NVIDIA Sortformer on overlapping 8-minute chunks, stitched with embedding clustering so labels stay consistent over hours of audio.

The product side is Next.js on Vercel: Firebase auth, Stripe credit packs, live transcript streaming into the UI, Claude-powered summaries and chat, 25 languages, and a free tier that runs Whisper entirely in the visitor's browser.

Most cycle-tracking apps predict the next period from the median of past cycles, which fails for anyone with irregular ones. At Bellabeat, I developed the prediction model behind the IVY Period & Pregnancy Tracker: one multi-task transformer that jointly predicts period start, period length, ovulation status, and days to ovulation from cycle history, logged symptoms, and wearable signals such as heart rate and skin temperature.

The results are described in a published article on the tracker's site. I took the model to production on-device, converting PyTorch to Core ML (iOS) and ONNX (Android) so predictions run offline, and health data stays on the phone. I built the data pipelines, monitoring, and governance on GCP.

RSNA 2024 asked competitors to grade spinal stenosis from MRI: 25 targets per patient across five conditions, five vertebral levels, and three severity grades. With one teammate, I finished 15th of 1,874 teams (top 1%).

Our pipeline had three stages. U-Net ensembles first locate the anatomical keypoints on each MRI series. Then, instead of training a model to tell left from right or assign vertebral levels, we computed that geometrically from DICOM metadata, which maps every pixel into 3D patient space. Finally, a shared CNN with three specialized heads grades the severity of crops around each keypoint, weighted to match the competition metric.

We validated by re-implementing the official metric locally over a 5-fold cross-validation, and shipped the whole thing as a fully offline Kaggle notebook.