Linearizable key-value store · Go

Raft from scratch: PreVote, ReadIndex linearizable reads, persist-before-send ordering, snapshots, a durable WAL, and a gRPC runtime.

A deterministic harness replayed partitions, crashes, message loss, delays and snapshot churn across 2.9M client operations: zero safety violations, and zero Illegal verdicts from the linearizability checker. The first bug it caught was in my own election-safety check, not in the consensus code.

Crash-safe LSM storage engine · Rust

A CRC-framed write-ahead log with group commit, memtables, block SSTables with bloom filters, and size-tiered compaction. unsafe is forbidden at the crate root, so the engine has no unsafe blocks by construction.

Power loss is replayed at 330 deterministic crash points (2,640 executions) plus 160 property-based schedules: zero acknowledged-write loss, and a ~29x gain from group commit. Benchmarked against sled, which wins every matched comparison on this host. Reported rather than omitted.

Vector search engine · Rust

An HNSW proximity graph and an IVF-PQ compressed index over hand-written AVX-512 distance kernels with runtime feature dispatch, for embeddings and RAG retrieval.

On 50,000 128-dimensional vectors on one 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 9–33x across the 90%+ recall band. The honest catch: that baseline runs in parallel across all 48 vCPU while the search runs on one, so the two sides of the ratio aren't core-for-core.

CCD-data-archive

Four merged pull requests to the UMD Observatory data archive, in Python, Flask and SQLite, 50,000+ records.

lambeq #259

One pull request open upstream, adding a LAMBEQ_MODELS_URL override for model downloads.

Quantum NLP (FIRE)

A four-person research project, co-built and open-sourced, in DisCoPy, Qiskit, pytket and JAX. My part was dataset preparation, experiment integration, and documentation.

_{The ML, NLP and statistics work is on the portfolio ↗}