🎓 BS Computer Science & Computational Physics @ UT Austin

I build systems at the intersection of low-level performance, machine learning, and computational physics.

• ⚛️ qmc-pricer Quantum Amplitude Estimation for derivative pricing, built from scratch — including the simulator — and benchmarked against classical Monte Carlo and quasi-Monte Carlo.

  • The quadratic speedup is real, and it is not enough. IQAE converges at N^-1.01 against Monte Carlo's N^-0.54 with non-overlapping confidence intervals — exactly what the literature promises. Scrambled-Sobol quasi-Monte Carlo reaches the same rate classically and beats every quantum method by two orders of magnitude in absolute error at every budget tested.
  • The simulator is mine. A from-scratch statevector engine in Rust — memory-bandwidth-aware kernels, gate fusion, rayon — hitting 94% of this machine's measured 270 GB/s sustained bandwidth at 26 qubits, slightly faster than Qiskit Aer. The denominator is benchmarked, not a spec sheet.
  • Two independent backends. A Rust engine (hand-written index arithmetic over a flat buffer) and a JAX engine (tensor contractions) sharing no code, agreeing with each other and Qiskit to 4e-16. A bug would have to be reproduced independently, in two languages, by two different algorithms, to survive.
  • What it would actually cost: ~9.2e12 T gates and 61 logical qubits to price one option to 1e-4 — and the polynomial state preparation that would close that gap is precisely what Herbert's critique targets.

• 🔬 Preimage Amortized prompt inversion, evaluated functionally. Train a model that reads arbitrary text and emits a prompt making a frozen target model reproduce that text's distribution — judged by whether the regeneration matches, not by string-matching a "true" prompt (which, for arbitrary text, doesn't exist). At its core: the inverse problem for a stochastic simulator, i.e. amortized Bayesian inference.

  • Amortization beats search: the reward-optimized inverter reaches 93% of best-of-8 search quality in a single call (0.681 vs 0.732 embedding similarity) — above a 0.637 zero-shot bar and a 0.359 retrieval floor.
  • Reward hacking, measured and priced out: ~24% of sampled candidate prompts attempt verbatim-copy strategies, with functional score correlating +0.56 with target overlap absent the guard. The n-gram guard acts as a Bayesian prior term.
  • Built so the numbers can't be artifacts: template-family-disjoint holdouts, pre-split MinHash dedup, provenance-split memorization controls separating reachability from recall, and metric sanity gates validated before any training.
  • Status: complete config-driven pipeline, 52 tests, validated end-to-end at dev tier (135M stand-in). Full-scale 8B runs are config changes pending GPU allocation.

• 🧠 Lohalloc An intelligent, machine-learning-driven memory allocator that bridges ML and low-level systems architecture. A UCB1 multi-armed bandit learns per-call-site allocation topology during training, then freezes into a CHD minimal perfect hash table for O(1) routing at inference.

  • A specialist, honestly priced: wins mixed-size/lifetime workloads by up to 6× on speed and 4–16× on peak RSS — and loses uniform small-object churn by 1.5–1.8×. Certified on bare-metal AWS Graviton across C/C++/Rust with Mann–Whitney U significance testing over raw per-run samples.
  • The bug worth reading about: the bandit kept under-selecting the bump arena. The cause wasn't the allocator — a 48-byte training header was landing on the arena's cold target cache line, paying exactly the cache-miss cost the arena existed to avoid. The reward signal was measuring the instrument, not the system.
  • Live Telemetry Dashboard: a custom real-time GUI (LOHA // ALLOC) monitoring heap maps, fragmentation, ops/sec, and latency — zero-overhead in production builds (feature-gated; verified via nm).

• 🎛️ CRATEDIG A real-time terminal sampler and sequencer in C++20, built because sample-based workflows are the one thing GarageBand handles poorly and the free alternatives are paywalled.

  • Hard real-time, enforced rather than intended: zero allocations, locks, syscalls, or exceptions in the audio callback — three lanes over lock-free SPSC rings with a janitor thread that owns all destruction, checked by global operator new replacement against a thread-local depth counter and a ThreadSanitizer-clean suite.
  • Bit-exact determinism: identical output across block sizes, run-to-run, and live-vs-offline, verified by committed golden hashes with a documented FMA-portability policy. Playback position is 32.32 fixed point, not float, so block-size invariance holds by construction rather than by luck.
  • Negative-control testing: every assertion is validated by deliberately breaking the behavior it claims to catch. It found a full suite that passed with bus routing deleted entirely.
  • Status: M5.7 of a documented 9-milestone roadmap. Chops, sequences, mixes. Not yet: recording, export, plugin hosting.

• 🦀 Iron Clad A high-performance Unix shell in Rust, built to find out how much of a shell's latency is the shell's own fault. Unified my interests in core systems engineering (architecting the shell itself) and cloud infrastructure (AWS for rigorous benchmarking and validation).

  • Performance, on parity-verified workloads: geometric mean 1.21× over 15 cases vs bash — 1.84× REPL dispatch (45.6% lower latency), 1.27× spawn-and-reap at 250 concurrent jobs (21.2%), 1.24× pipeline at 16 stages (19.4%), 1.14× cold start (12.1%). Via posix_spawn-only execution (zero fork() calls, no CoW page-table cost), zero-copy fd passing (libc::dup2), and a borrowing parser over a recycled input buffer — 3 heap allocations for a simple command, pinned by a counting-allocator test.
  • Where it doesn't win, stated: parse complexity 3 sits at 0.99×, still a hair behind bash. Two parse cases were worse — 0.80×/0.83×, from spawning /bin/echo on every echo — until an echo builtin moved them to 1.15× and 0.99×.
  • Measured properly: every benchmark pair is gated on a semantic parity check (stdout, exit status, and failure diagnostics must agree before timing), since Iron Clad implements a POSIX subset and an unsupported-syntax case would otherwise time bash doing work it skips. hyperfine -N at p50/p90/p99 on AWS Graviton bare metal via a zero-touch Terraform + Packer pipeline with automated isolcpus core isolation. Cold start n=100; other cases n=10, so p50 is the reportable statistic.

• 📦 WarrenDB — a distributed, erasure-coded, self-healing object store in Go, built as a controlled experiment in how storage systems should spend their repair I/O. ~1,200 seeded runs show corruption-aware scrub scheduling cuts unrecoverable data loss 54% at equal I/O budget, and a deterministic harness collapses 30 days of cluster failure into a 14-second bit-for-bit reproducible run over the real production code path.
• 📈 Proper Time — reproducing stochastic subordination (Clark 1973; Ané–Geman 2000) on 32M BTCUSDT tick trades: excess kurtosis collapses ~32 → ~2 moving from calendar to event time, with two negative results reported alongside it and an original dτ/dt "clock-rate" chart showing information intensity spiking ~100× around scheduled shocks.

• Core Interests: Systems, Machine Learning, Mathematics, Quant Research, and Quantum Computing.
• Currently Learning: Compilers, GPU Software, Distributed Systems, Networking, and Market Microstructure.

• Pronouns: he/him
• Fun Fact: I like coffee, basketball, and weightlifting