Work
Case studies and research, with the limits stated.
Every project here says what was done, what was measured, and what was not claimed. Client names and confidential metrics stay out of the public record.
How to read this page
Each card states the context and the metrics that can be published. The detail page adds a "What I did not claim" section, because the boundary of a project is part of its evidence. If a claim matters to your decision, ask for the walkthrough.
01 / Automation and delivery
Workflows and systems delivered in a real environment.
Assistant, agent-platform and deployment-workflow projects, with client names withheld where confidentiality applies.
An internal AI assistant for a US pediatric hospital
Owned the AI and application layer, and served as the hospital’s primary technical point of contact for the AI and application scope.
A private agent platform with tools on NVIDIA DGX Spark
Deployed NVIDIA AIQ on NVIDIA DGX Spark and extended it with tools and agent capabilities, on customer-owned hardware.
An agentic workflow that deploys a repository to local, Docker or Azure
Built during an AI internship: a platform that takes a repository and agentically deploys it to a local environment, Docker or Azure.
02 / Research and independent work
Research that had to survive a measurement.
Thesis work, medical imaging and retrieval research. Research is not a client deployment, and these pages say so.
MTF-Qwen: a 0.6B model against six forecasting baselines
B.Sc. thesis: fully fine-tuned Qwen3-0.6B for multi-horizon cryptocurrency forecasting, with a custom loss and a controlled benchmark against six deep-forecasting baselines.
Lung cancer radiogenomics: 92% recall, 80% fewer false positives
Combined NVIDIA VISTA-3D segmentation with a custom ResNet-18 classifier on LIDC-IDRI to predict EGFR/KRAS biomarkers from lung CT.
Token Sniper: minimal retrieval over a code knowledge graph
Minimal sufficient retrieval for repository-level code intelligence, using a Variational Graph Information Bottleneck over a code knowledge graph.
A trajectory-aware Transformer built in PyTorch
A self-directed architecture study: a trajectory-aware Transformer for multi-robot path planning, with positional encoding written from scratch.
Next step
Want the walkthrough?
Some of the most useful evidence is in private repositories and confidential delivery. I can walk you through the code, the evaluation protocol or the architecture on a call.
Prefer email? boroicamarius01@gmail.com