I study computer science at Michigan and like building things people actually use. Lately that’s meant cardiac ML research at Michigan Medicine and the sales platform Kareem Mart runs on.
I build things people use.
a sales platform, $500K+ through itI look for patterns in heartbeats.
cardiac ML at Michigan MedicineI make systems that don’t fall over.
Go, replication, MapReduceMatched 18,000+ ECGs to echocardiograms for 5,400+ patients with left bundle branch block, on a HIPAA-compliant HPC cluster. Wrote a multi-scale Haar wavelet transform from scratch in NumPy, tested 1,920 features with FDR correction, and isolated 12 bands that differ between reduced and preserved heart function.
Linked 1.9 million ECGs with 303,000 echocardiograms, narrowed them to a 592-patient cohort in SQL, and trained an XGBoost model that predicts who responds to cardiac resynchronization therapy (0.71 mean AUROC with patient-grouped cross-validation, pre-op data only). It won a Blue Ribbon at the spring symposium.
Fixed bugs in a webhook delivery system on AWS Lambda and shipped a new function through GitHub Actions.
Built them a sales platform in Next.js and Supabase and wired it to Zoho Inventory, so 6,000+ products stay in sync and there’s 60% less invoice typing by hand. 26 people use it to manage inventory, customers, and orders, and $500K+ in orders has gone through it so far.
Led architecture on team projects, built a digital art gallery app, and helped newer developers get comfortable with Swift.
One LLM call doing pre-op triage scored 70.7%. So I split the job: the model only extracts quoted facts, and Python applies every rule. That took the eval score to 99.5%. Each quote gets checked against its source document, and anything unverifiable goes to a human, with 0 false clearances on 50 sample and 15 held-out cases.
Primary/backup replication in Go. A view service promotes the backup when the primary goes quiet and never lets two primaries exist, even across a network partition. The catch: no mutexes allowed, only goroutines and unbuffered channels.
Time-series vitals and labs for 16,000+ patients, an RBF-kernel SVM weighted for the class imbalance, and a lot of cross-validation on C and γ. It landed at 0.86 AUROC (bootstrapped CI 0.81–0.91).
First the MapReduce framework: a manager, workers, and UDP heartbeats to notice when one dies. Then I used it to index 100,000+ web pages, ranked with tf-idf mixed with PageRank.
Python for anything with data, and C++ when I want to know exactly what the machine is doing. Plus SQL, PyTorch, scikit-learn, XGBoost, Postgres, AWS Lambda, and a research HPC cluster I know too well. Lately I love exploring what Claude can do. Claude Code is part of how I work, mostly on the platform and front-end work like this site.