Every project links to its code. Any number here comes from a benchmark committed in that repo, so you can re-run it yourself.
FIG. 01 · parallax-kv
DISTRIBUTED SYSTEMS · GO
A three-node Raft key-value store built from scratch: PreVote, ReadIndex, persist-before-send ordering, snapshots, a durable WAL, a gRPC runtime.
A deterministic stress harness ran 2.9 million client operations through partitions, crashes, message loss, delays, and snapshot churn. Zero safety violations, and zero Illegal verdicts from the linearizability checker.
Inspect the repository (opens in new tab)docs/BUG_LEDGER.md (opens in new tab)
Specifications
Notes & limits
The first bug the harness caught was in my own election-safety check, not in consensus. Not built: production snapshot scheduling or chunked streaming.
3 nodes · soak: 6 simulated clients, virtual time · throughput: localhost, c=8, 4 KiB fsync p50 ≈ 0.89 ms · measured 2026-07-13
FIG. 02 · accretion-db
STORAGE ENGINE · RUST
An embeddable LSM-tree storage engine: CRC-framed WAL, group commit, memtables, block SSTables, bloom filters, size-tiered compaction. The crate root forbids unsafe code, so the engine has no unsafe blocks by construction.
Fault-injection and property tests confirm that acknowledged writes survive simulated crashes. Benchmarked against sled, which wins every matched comparison on this host. Reported rather than omitted.
Inspect the repository (opens in new tab)benchmarks/RESULTS.md (opens in new tab)
Specifications
Notes & limits
The ~29x group-commit figure is WAL-bound. Compaction runs synchronously under the write lock, which caps full-engine throughput.
16-byte keys / 100-byte values · ext4 on NVMe · Xeon 8488C · measured 2026-07-13
FIG. 03 · lodestone
VECTOR SEARCH · RUST
A vector search engine built from scratch in Rust for embeddings and RAG retrieval. An HNSW proximity graph and an IVF-PQ compressed index sit over hand-written AVX-512 distance kernels with runtime feature dispatch.
On 50,000 128-dimensional vectors on a single core, HNSW reached 0.976 recall@10 at about 31,800 queries per second, about 33x the exact brute-force search at that recall, and 9 to 33x across the 90%+ recall band. The baseline is the honest catch: it runs in parallel across all 48 vCPU while the search runs on one, so the two sides of that ratio are not core-for-core.
Inspect the repository (opens in new tab)benchmarks/raw/bench_50k_128d.txt (opens in new tab)
Specifications
Notes & limits
IVF-PQ trades accuracy for footprint: 0.975 recall@10 at 16x memory compression. Single machine, single core. Not built: distributed sharding.
50,000 x 128-d · k=10 · search on 1 core, brute-force baseline on 48 · Xeon 8488C · measured 2026-07-16
DistilBERT with a gradient-reversal DANN for unsupervised domain adaptation across Yelp, Amazon, Twitter, and Reddit. Labeled source-domain examples train the sentiment head; unlabeled target-domain text trains the domain head.
The committed CSV is a smoke-scale artifact, so this page quotes no accuracy numbers. The full figures live in the linked repo's team report.
Four-person research project, co-built and open-sourced. My role was dataset preparation, experiment integration, and project documentation.
The repository compares optimizer and ansatz variants in exact noiseless simulation with documented methodology.
Hands-on generative-AI notebooks. The core is a retrieval-augmented generation (RAG) pipeline built stage by stage in LangChain, from tokenizers through chunking and embeddings to retrieval and QA chat. Alongside it: Streamlit/Gradio LLM apps and GAN/VAE architecture studies in PyTorch.
Each notebook isolates one idea.
Predicts a professor's average star rating on PlanetTerp from the text of student reviews (VADER sentiment plus review count), without ever using the star ratings as features.
Two text-derived features in a plain linear regression explain about 79% of the variance in ratings.
Cleaning and statistical analysis of a multi-wave survey dataset: a reusable cleaning function, then chi-squared hypothesis tests with Bonferroni correction probing what actually shifts people's answers to moral-judgment questions.
Does swapping the genders in a moral-judgment scenario change how people judge it? Chi-squared hypothesis testing on gender-swapped survey scenarios, with the method written down.
A statistical study of attendance against local weather: per-group detrending, WMO/NWS rain classification, Welch's t-tests, then a rain-plus-wind regression.
The page you're reading: a multi-page React + TypeScript + Vite site on GitHub Pages. It prerenders to static HTML, works without JavaScript, and won't deploy if a fact check or Playwright assertion fails.
I'm looking for Summer 2027 software engineering internships and new-grad roles. Email is the fastest way to reach me, and I can send a full resume on request.