Mathematics BS, Montana State University, May 2026 — completed in three years, working through college. Independent researcher at the intersection of reinforcement learning, mechanistic interpretability, and AI safety — mostly on whether hidden behavior can be planted in a model, and whether it can be caught.
A Cartridge is a trained KV cache: a small, shippable artifact that installs knowledge or behavior into a model without touching its weights — so it's a plausible future supply-chain object. This asks whether a cartridge that does something genuinely useful can also carry a hidden trigger, and whether a defender can detect it. Used as a cheap, reversible model organism: no fine-tuning run, no weight diff, trainable in minutes on one consumer GPU. Run as two autonomous multi-day campaigns on rented GPUs, scoped and budgeted end to end (~$7.50 total).
Key result: Yes, if the useful function and the trigger are trained jointly — naive concatenation fails by destructive interference, independently replicating CAS. The resulting backdoor is concept-keyed, so it over-fires a surrounding region that an α-sweep shows is a magnitude-tuned zone, not a scale-invariant cone, and keyed more on spelling than meaning. Then the detector race: the backdoor is opaque to static KV extraction and invisible to a pre-output activation monitor even when triggered, but caught 4/4 by a response-position monitor that also names the payload — and faithfully flags an accidental leak that was never planted. The detection window is concurrent with firing, not before it.
PyTorch HuggingFace Transformers PEFT mechanistic interpretability activation steering model organisms
Empirical study of how Arditi et al.'s (NeurIPS 2024) "refusal lives in a single direction" finding transfers across model families. Replicated cleanly on Qwen 2.5 1.5B (10/10 coherent jailbreak, N=12 with explicit coherence checks), then mapped a 5x3 block-by-direction matrix across Gemma 2 2B, Gemma 3 1B, and Gemma 4 E2B to test single-layer ablation. Cross-Gemma results are dramatically asymmetric — Gemma 3 is the architectural outlier, not the rule.
Key result: Direct parameter inspection found that Gemma 3's post_attention_layernorm and post_feedforward_layernorm gains are 5–30x larger than Gemma 2, Gemma 4, or Qwen — amplifying per-block residual perturbations and making single-layer interventions disproportionately effective. Gemma 4 corrected the calibration. Random-direction control rules out the noise-injection alternative. Inference-only on a 12 GB consumer GPU.
PyTorch HuggingFace Transformers mechanistic interpretability forward hooks matplotlib
A PPO agent for the roguelike poker deckbuilder Balatro — eight architecture versions, ~366 hours of compute. Started by training against the live game through a Lua mod with file/socket IPC (V1–V3), then pivoted to a purpose-built Python simulation (164 jokers, consumables, boss blinds) with a 496-test regression suite, for a ~12,500x throughput speedup. The V7 agent uses a hierarchical action space — an intent head over play/discard/use plus a learned card-selection head — over a 434-dim observation at ~2.5M parameters, peaking at a 2.35% solo win rate and reaching ante 9 reliably.
Key result: The ~2% ceiling is capacity-insensitive. Six reward-shaping retunes, a 5.5x network scale-up, and four self-play variants all hit the same plateau — the bottleneck is the exploration mechanism, not network size. Concluded with a principled next axis: MCTS with a neural policy/value prior, not more model-free PPO. An earlier version's headline win rate was invalidated by a self-audit — fixed-seed memorization plus a miscoded joker — and is reported as retracted rather than quietly dropped.
Python PyTorch Gymnasium Lua PPO GAE multiprocessing
A certificate-first automated campaign against open problems in quasigroup, loop, and semigroup theory. Encodes candidate conjectures for ATP provers (Vampire, Twee), SAT solvers, and GAP, then runs them under a verification protocol designed so that nothing counts as a result without a machine-checkable certificate — an explicit guard against plausible-looking output that doesn't hold up.
Key result: Re-running the QPTP problem library with current provers showed that problems recorded as hard in the 2008–2010 studies now fall quickly — JKVxx_2 solves in 23s under Twee and 206s under Vampire. Corresponded with the library's authors (Stanovský, Kinyon) to confirm the finding and identify which problems remain genuinely open.
automated theorem proving Vampire Twee SAT GAP quasigroup/loop theory
Comparative evaluation of multi-signal pandemic early-warning, submitted to Apart Research's AIxBio Hackathon (Track 2) and externally reviewed. Tests whether four surveillance signal types — wastewater PCR, Google Trends, Wikipedia pageviews, and CDC syndromic data — keep calibrated relationships with clinical ground truth across a pathogen's transition from emerging to endemic, using COVID-19 as the subject and influenza as a controlled comparison.
Key result: Attention-based signals show 5–23x variance compression after the first major COVID-19 wave but none across flu seasons — attention decay is an emerging-disease novelty-cycle phenomenon, not a property of the signal type. Wastewater is the only signal type that holds calibration across the full lifecycle. Includes an honest negative result on LLM-conversation surveillance.
Python pandas HuggingFace Transformers time-series anomaly detection matplotlib
Systematic comparison of four ML model families for predicting ICU mortality on MIMIC-IV (74,829 patients). Feature-engineered XGBoost (AUROC 0.9565) significantly outperforms Neural ODEs (0.9039) and clinical language models (0.8809), with all differences confirmed by bootstrap CIs and DeLong significance tests. Investigated whether clinical text models exploit code status documentation (CMO/DNR) as a confound using a novel multi-task ClinicalBERT framework.
Key result: Structured time-series features dominate text for short-horizon ICU mortality. The interesting question is not which model wins, but what each modality actually captures.
Python PyTorch torchdiffeq HuggingFace XGBoost PEFT SLURM/HPC
Live Jekyll site of long-form reviews on AI safety, philosophy of science, and epistemology — Taleb, Christian, Callard, Hao, and others. Each review is its own argument, not a summary: what the book changed, where it falls short, and how it connects to alignment or my own research. Source: taggarttufte/book-reviews.
Jekyll GitHub Pages kramdown MathJax
Multi-Voice Audiobook Generator End-to-end EPUB → multi-voice MP3 audiobook pipeline using xAI's Grok TTS. Heuristic dialogue attribution with confidence scoring, gender-matched voice mapping, per-segment caching, and a bundled Flask web player with variable speed, bookmarks, and paragraph-synced read-along. 7 books rendered to date; ~70× cheaper to run than ElevenLabs at comparable quality.
Ski Resort Pricing Analysis Predictive modeling of season pass prices across 500+ resorts using multi-variable regression and Cholesky decomposition. Built from numerical foundations rather than black-box libraries.
SVD Image Compression Image compression via Singular Value Decomposition. 17x compression at rank-50 with less than 6% reconstruction error.
Python · PyTorch · NumPy · pandas · scikit-learn · HuggingFace Transformers · PEFT · Gymnasium · Flask · MATLAB · XGBoost · Git · SLURM · rented-GPU workflows (Vast, RunPod)
Strong foundation in numerical linear algebra, real analysis, and statistical theory. Experience with HPC (MSU Tempest, NVIDIA A40) and rented single-GPU cloud instances for training and interpretability work.
Most of what I do is small-scale and carefully controlled rather than large-scale: norm-matched random-direction controls, held-out baselines, multi-seed checks where variance matters, and null results reported as nulls. Two of the projects above include a public retraction of my own earlier result after a self-audit. Interested in AI safety and alignment — particularly deception, backdoors, and whether hidden behavior can be detected.