I build and scale AI-enabled biological systems. My research focuses on the intersection of protein language models, high-dimensional statistics, and high-performance computing.

The recovery threshold under distribution shift.

Protein Language Models · Distribution Shift

A three-layer classifier over 1280-D ESM-2 embeddings, trained on one protein population and adapted batch by batch toward a shifted one. Swapping the Gaussian-mixture simulator for real UniProt embeddings under a Bacteria→Archaea taxonomic shift surfaces the negative-transfer dip the synthetic construction structurally forbids: target F1 falls below its pre-adaptation baseline at low out-of-distribution pool fraction and recovers only once that fraction reaches roughly one half. The recovery threshold sits at r ≈ 0.5.

GitHub

Online Bayesian drift tracking for QEC decoders.

Quantum Error Correction · Sequential Monte Carlo

A surface-code matching decoder decays as hardware noise drifts away from its one-time calibration. qecdrift is a bootstrap particle filter over two channel scales that consumes the live syndrome stream and re-primes the decoder between batches. Uniform drift leaves matching invariant, so only the relative gate/measurement split is tracked, and an exact collapse of the composite likelihood onto its interpolation grid holds a million particles at ~2 ms per batch on an H100. It cuts post-event logical error 10-14% in 4M-shot sweeps, and a strictly causal replay of all 248 Willow configurations certifies most of the device stable while resolving one readout ramp to 1.81×.

GitHub(private)

Data Diversity and Sequence Length: Key Levers for Powerful Biological AI.

Bioinformatics · Foundation Models

An invited synthesis, with Adrienne Hoarfrost, of the two levers that set the transfer-learning ceiling of biological foundation models: the taxonomic and functional biodiversity of the pretraining corpus, and the maximum sequence context the model can ingest, with architecture as a coupled third variable. The tension is structural: biodiversity concentrates in short metagenomic reads of 100-300 base pairs while the dependencies that reward long context span kilobases to megabases, so a single corpus cannot maximize both. Adding raw data or raw context without a matching architecture yields diminishing returns; the largest untapped gains lie in architectures that fuse short diverse reads with long assembled context. It treats the dual-use biosecurity exposure as intrinsic to generative capability and names the controls. NATO Science for Peace and Security book chapter, Springer 2026.

A nightly portfolio desk under coded risk vetoes.

Quantitative Finance · Regime-Switching Monte Carlo

A two-stage SLURM pipeline that, each market night, fits a 2-state Markov-switching mixture per asset, simulates forward paths by circular block bootstrap over the asset's own return history, and reduces each to P(up in 6m), CVaR, and skew. A LangGraph DAG under a Llama-3.3-70B server walks a macro strategist, per-ticker news sentiment, a portfolio manager, and a risk officer whose position, trade, and cash limits fire in code, not prompts. No broker API exists anywhere in the pipeline: the desk grades its own closed calls against SPY and cannot execute.

GitHub

Hiring signals from organizational semantic drift.

Labor-Market Signals · Gaussian HMM

A nightly pipeline embeds each organization's authored public documents, grants, preprints, and filings, into a per-√day semantic-shift series, so a two-day pivot in terminology space outranks the same drift spread over eight months. A 2-state Gaussian HMM separates steady voices from teams pivoting hard, regime-switched Monte Carlo yields a next-quarter escalation probability, and a Hawkes process flags organizations hiring hot even when net headcount is flat. A LangGraph DAG under Qwen3-Next-80B drafts evidence-cited outreach memos; a human reviews and sends everything, and no send path exists.

GitHub

MDSmooth: turning-point key frames from MD trajectories.

A ChimeraX bundle that reduces a molecular-dynamics trajectory to its turning-point frames by zero-phase Butterworth filtering a single collective variable (RMSD, PC1, tICA IC1, or dihedral PCA), then morphs between the filtered extrema along a corkscrew screw path. The one cutoff is solved by bisection to hit a target key-frame count. Every signal reports a cosine content that calls out when the slow mode is undersampled diffusion masquerading as a transition, so the tool refuses to dress up frames the simulation never sampled.

GitHub/ChimeraX Toolshed

Identifiability limits of a two-scale biofilter model.

A packed-bed biofilter modeled as a biofilm reaction-diffusion BVP coupled to a column advection-dispersion-reaction BVP with Monod kinetics, calibrated to 2006 bench data by affine-invariant MCMC with one nested BVP solve per likelihood. Simulation-based calibration passes for all twelve parameters, yet the posterior design volume spans three orders of magnitude: the bench runs entirely below the half-saturation constant, so only the lumped rate R_max/K_s is identifiable. Refitting the one first-order rate the data actually constrain collapses the interval to a factor of 2.3 and reproduces the original point design.

GitHub

Bioprocess Optimization via Evolutionary Algorithms.

Engineered a stiff fed-batch bioreactor simulator for L-asparaginase production in metabolically engineered E. coli, then drove it with two gradient-free optimizers. A particle swarm (PySwarm) tunes the PID temperature controller; a genetic algorithm (DEAP) evolves the reactor design for maximum profit. The model is blunt about its own economics: the process pays only once acetate overflow is engineered out of the strain.

GitHub

Applied Research Scientist - US Army Corps of Engineers.

Contract work focused on applied research and development within nanotechnology, delivering highly technical, scalable solutions tailored to defense-oriented scientific applications.

Details withheld: work performed under federal nondisclosure. No deep-dive available.