AI infrastructure
Reproduction artifact for the PEARC '26 paper 'LLMs or Naive Bayes? Old Gems or New Ways'. Complement Naive Bayes vs zero/few-shot LLMs across four model families and a 37x scale range, with a Helm operator that picks the model by policy and Prometheus metrics you can verify.
Prometheus exporter for Qualcomm Cloud AI 100 accelerators. Scrapes qaic-util every 15 seconds and exposes 20+ per-QID metrics (power, temperature, DRAM, NSP utilization, inference load). Injected as a ConfigMap, no custom image needed.
The 30-minute AI infrastructure talk from the 7th National Research Platform Workshop (2026, UC San Diego / SDSC), built with Slidev.
Archive
Single-image facial super-resolution with GANs, built for CS 584 at Illinois Tech.
Grayscale image colorization with CNNs and GANs, built for CS 512 Computer Vision at Illinois Tech.
Naive Bayes classifier for sentiment analysis on Amazon reviews, built from scratch for CS 481 at Illinois Tech. The seed of the benchmarking work that became PEARC '26.
Kubernetes & platforms
Gives plain JupyterHub the features that usually belong to Coder: SSH into user server pods over HTTPS, wildcard subdomain port forwarding, and one-click VS Code into a pod's workspace. Deploys as an overlay using only published extension points, no changes to core JupyterHub.
Multi-hub management for JupyterHub on Kubernetes: a hub that provisions and manages many JupyterHub Helm releases in separate namespaces, with a Kubernetes-based HubSpawner, RBAC, an authenticator, and a REST API for programmatic control.
Java eBPF bytecode generator plus PCAP tools: custom packet filter rules compiled to eBPF, a PCAP reader/writer, packet filtering and manipulation, and an IPv4 packet generator.
Network data & measurement
Just Another Interpreter Like Yours: a Java transpiler that converts natural-language rules (NRL) into BPF packet filters. Parser plus code generator, stateful rules, manual packet manipulation, evaluated against PCAP files, applied with a C toolchain. Tell your filter what you want in English.
Minimal sFlow collector for Kubernetes: listens on UDP/6343 behind a ClusterIP service, decodes samples to JSON (sflowtool -J) or PCAP (-t), and self-tests with a synthetic packet. The practical way to get real packet data on the National Research Platform.
Python packet generation with Scapy: build and customize packets across protocols and parameters, for testing network paths, switch behavior, and sFlow collection pipelines. The lab gear that gets used more than it gets documented.
Yes, it's named Naruto. It's actually NaRuTo, a Natural Rule Tool: an extended parser for the Natural Rule Language, letting you express model constraints in near-English and write mapping programs in Python or JavaScript, not just Java. The grammar half of the Jaily story.