Projects

It Started in Bed I grabbed my laptop at midnight, propped up on pillows, and started typing. No grand plan. No business case. Just an itch that wouldn’t go away: why does self-hosted artifact management cost so much and lock you in so hard? I’d seen teams paying six figures a year for JFrog Artifactory. I’d watched Nexus deployments turn into maintenance nightmares. And I kept thinking – it’s 2026, we have Rust, we have modern tooling, we have AI that can actually help write code. What if someone just… built the thing? Open source, MIT licensed, no enterprise tier, no feature gates. ...

A 6-part series investigating why Docker containers use 20-30GB more memory than native execution on Grace Blackwell’s unified memory architecture. Overview When YouTube reviewers complained about DGX Spark performance, they blamed the hardware. I dug deeper and found the real culprit: Docker’s cgroups double-counting unified memory. Key Findings Container overhead: 20-30 GB more memory usage in Docker KV cache reduction: 40-63% less KV cache in containers (1.7-2.7x reduction) Performance: Identical throughput - no speed penalty Root cause: Docker’s cgroups double-count unified memory on Grace Blackwell Solution: Use native execution for large models on unified memory architectures The Series The Mystery - YouTubers blamed NVIDIA hardware without technical analysis MPI and Chroot Nightmare - Setting up proper test environments The Unified Memory Revelation - Why Docker’s cgroups double-count unified memory The Data: 60 Runs Don’t Lie - Comprehensive benchmark results across 3 models What I Learned (And What’s Next) - Conclusions and Phase 2 preview KV Cache Deep Dive - The 2x reduction mystery explained Related Resources Benchmark Code & Data: benchmark-spark Interactive Results: brandonrc.github.io/benchmark-spark Impact This investigation demonstrates the importance of understanding the full stack - from hardware architecture to kernel subsystems to container runtimes. What looked like a hardware problem was actually a software architecture mismatch. ...

Final Test Flight An autonomous flight system built around NVIDIA’s Jetson Nano running YOLOv3 for real-time object detection and navigation. Project Overview During my time at the University of Minnesota Duluth, I built an autonomous octocopter that combined embedded AI with flight control systems. The goal was to create a drone capable of autonomous navigation using real-time computer vision. ...