I build full-stack products, ship ML systems, and care about code that holds up under real conditions.
Work & Research
Scaled a self-built vacation rental analytics SaaS to 26k+ bookings and $5M+ in tracked revenue across 80 properties. Cut dashboard loads to 50–100ms by consolidating DB calls (26 to 8 per view) and adding targeted cache invalidation.
Saved ~$40k/year in reporting labor by automating ETL pipelines across 12 branches. Cut weekly reporting cycles by 85%. Reduced unplanned equipment downtime by 22% via predictive maintenance models deployed on 800+ industrial compressors.
Grew MadHacks from 257 to 419 attendees and doubled project submissions over two years, the largest in-person event ever run by UW-Madison CS students. Hold twice-weekly office hours advising peers on projects, research, and careers.
Benchmarked CDVAE-style generative autoencoders on 100K+ crystal structures for inverse property prediction, part of a program featured at the White House Materials Genome Initiative. Cut experiment turnaround from weeks to days across 200+ HTCondor compute nodes.
Selected Projects
Network Topology Mapper
C program that parallel traces multiple hosts and renders the merged path tree live. Hand-rolled raw IPv4/IPv6 ICMP sockets, encoding target and TTL into sequence numbers to correlate replies across concurrent traces without a packet-capture library.
Handwritten Code Compiler
MadHacks 2023Reads handwritten Java from a photograph and executes it. Node.js backend pipes Google Cloud Vision OCR through Prettier to normalize whitespace, then invokes javac at runtime. Built in 24 hours with no prior OCR experience.
Photonic QGAN
IQuHACK 2024 · MITGAN trained on a photonic quantum computer to mimic a maximally entangled four-state target. Generator and discriminator implemented as Perceval linear-optical circuits with SLOS simulation; PyTorch handles gradients. Reached 45% fidelity in 24 hours across three optimization approaches.
More on GitHub