Exploring the space where systems think and biology inspires.

Leidos' satellite fleet needed to schedule competing observation requests against a shared set of ground assets — a classic multi-objective optimization problem. I applied dynamic programming and memoization in Python to prototype a scheduling engine, then rewrote the entire engine in Rust once the approach proved out, since the team's production microservice ecosystem demanded stricter reliability and lower latency than a Python service could guarantee at scale.

To keep scheduling decoupled from the rest of the satellite operations stack, I designed an event-driven data pipeline on Kafka with at-least-once delivery semantics, persisting results to MongoDB. That decoupling meant any one service could fail or redeploy without taking down the mission-critical scheduling path around it.

I containerized the team's test suite with Docker and ran it on Kubernetes, wiring the whole thing into GitHub Actions CI/CD. Writing pytest coverage across boundary conditions and deliberate fault injection meant regressions got caught in CI before they ever reached production.

Separately, I built a Claude Code MCP plugin from scratch on top of an institutional knowledge graph mapping the team's ontologies and service dependencies. It now serves roughly 30 engineers across a multi-team defense program, giving them a shared, queryable map of how services relate instead of relying on tribal knowledge.

The lab's spatial transcriptomics work spans single-cell RNA sequencing (scRNA-seq), next-generation sequencing (NGS), and spatial transcriptomics data, each of which suits different model architectures. I built an end-to-end ML pipeline in PyTorch and TensorFlow with an agentic layer that automatically selects and evaluates the right model — scVI, scANVI, scGen, or AmortizedLDA — for a given dataset, with every run tracked in Weights & Biases. Tuning that pipeline for scale cut GPU runtime from 80 minutes down to 27 on identical hardware, a roughly 300% reduction that matters when the lab is iterating across dozens of datasets.

Keeping up with clinical genomics literature was its own bottleneck, so I built TransferAgent to automate NLP-based literature synthesis and cross-paper validity scoring across roughly 500 clinical genomics papers. It cut review time by about 60%, and I worked directly with clinicians in the lab to make sure the synthesis reflected clinical relevance, not just statistical significance.

That pipeline work is now feeding into an ongoing clinical genomics publication targeting a peer-reviewed computational biology journal, where I've been validating model performance and pipeline reproducibility for glioblastoma gene therapy research across a 10-person interdisciplinary lab.

I built a skin cancer detection pipeline using OpenCV for image preprocessing and scikit-learn (with NumPy/Pandas handling the data wrangling) that reached 95% classification accuracy across roughly 70,000 dermoscopic images. That work became my first published paper, 'Early Skin Cancer Detection Improvement,' through Inspirit AI.

In parallel, I trained and validated a Random Forest classifier on miRNA sequences for cancer cell likelihood prediction, reaching 95% predictive accuracy at p < 0.05, validated with stratified k-fold cross-validation to make sure the result generalized rather than overfitting to one split of the data.

I held a dual role as tournament director and head coach at Princeton Racket Club: directing 8 regional tournaments end to end — 380+ match entries, every one starting on time — while also coaching 3 classes of 8-10 athletes spanning ages 8 to 65.

On the coaching side, I built data-driven, skill-adapted regimens for individual athletes rather than running one generic program for everyone. That approach guided 3 junior athletes to over 150 regional ranking points combined and helped 5 adult players climb a full 0.5 NTRP competitive level.

I founded an independent racket stringing business from scratch and scaled it to 45+ regular clients, servicing 12-20 rackets a month across two service tiers: a standard 3-4 day turnaround and a premium 2-day option for players who needed their racket back fast.

Growing the client base meant leaning on grassroots channels rather than paid marketing — word-of-mouth referrals, flyers at local parks, referral discounts through high school tennis teams, and an Instagram page showcasing the stringing craftsmanship itself.