UMD CS B.S./M.S. · GRADUATING MAY 2027 · AWS SDE INTERN, SUMMER 2026

Ivan Wang

I mostly build things to find out how they work. Outside of work I'm into mechanical keyboards, bouldering, and keeping up with tech.

Open to Summer 2027 internships and new-grad roles · relocating anywhere in the US

Three of the things I've built, with one measured result each. The names link to the code.

parallax-kv (opens in new tab)Raft key-value store2.9Mclient operations checkedreads and writes replayed under failure
accretion-db (opens in new tab)LSM storage engine878 µsfdatasync p50median durable-write flush
lodestone (opens in new tab)vector search engine0.976recall@10top-10 search accuracy

This summer I'm an SDE intern on EC2 Networking at AWS, where I built an agent that migrated seven production canaries from Scala to Java, cutting each one from about a week of work to about a day. Full timeline

Resume on request · ivanwang8989@gmail.com

Outside of work

A few things I spend my free time on.

Building keyboards

You can customize every part, so you can tailor a board to exactly what you or someone else needs, and picking the right part for the job is the fun of it.

~`!1@2#3$4%5^6&7*8(9)0_-+=deletedeltabQWERTYUIOP{[}]|\pg↑capsASDFGHJKL:;"'returnpg↓shiftZXCVBNM<,>.?/shiftendctrlcmdaltaltfn
65% layout, drawn to unit width: 1u alphas, 1.75u caps, 6.25u space, and the right-hand column an arrow cluster needs.

Bouldering

I climb for the community and the problem solving. A boulder can usually be done a lot of different ways, and which one costs you the least depends on your body, and working that out is my favorite part.

Bouldering

One boulder, three ways up.

Following the industry

I keep up with tech pretty closely, from new hardware to how governments are handling AI. I'm interested in where the industry is heading.

What I follow

hardware
launches and reviews
AI policy
regulation, worldwide
keyboards
switches and group buys
the industry
where it's heading

Featured engineering work

FIG. 03 DETAIL · lodestone · recall vs throughput

lodestone is a vector search engine I built in Rust. This is its benchmark: higher is more accurate, further right is faster, and each dot is one setting I measured.

recall 0.90 floor ·ef=16 falls below it 2K 5K 10K 20K 50K 1.00 0.95 0.90 0.80 0.70 0.60 0.50 THROUGHPUT (QUERIES / SECOND, LOG SCALE) RECALL (AT k=10, 0 TO 1) HNSW m=16 IVF-PQ nlist=256 m=32 oracle · exact brute force oracle · 971 q/s exact · 48 vCPU nprobe=32 nprobe=16 nprobe=8 nprobe=4 ef=256 nprobe=1 ef=128 ef=64 ef=32 ef=16 ef=32 · recall 0.9759 ·31,845 q/s TEST CONDITIONS n=50,000 · d=128 k=10 · 1 core avx512 · Xeon 8488C
FIG. 03 DETAIL, as measured values: ten samples, ordered by throughput, at the source file's own precision.
IndexSettingRecall at k=10Queries per second
IVF-PQnprobe=320.97471,229
IVF-PQnprobe=160.97472,305
IVF-PQnprobe=80.97484,110
IVF-PQnprobe=40.94096,128
HNSWef=2561.00008,898
IVF-PQnprobe=10.48509,797
HNSWef=1281.000013,514
HNSWef=640.996420,946
HNSWef=320.975931,845
HNSWef=160.899647,335
Baselineexact brute force (parallel, full corpus scan)exact971

readingef=32 · recall 0.9759 · 31,845 q/s · single coredrag across the plot · or press ← →tap a point · drag to pan · or step

50,000 x 128-d · k=10 · single core · Xeon 8488C · bench_50k_128d.txt (opens in new tab)

FIG. 01 · parallax-kv

DISTRIBUTED SYSTEMS · GO

parallax-kv (opens in new tab)

A three-node Raft key-value store built from scratch: PreVote, ReadIndex, persist-before-send ordering, snapshots, a durable WAL, a gRPC runtime.

GoRaftgRPCPorcupinedeterministic simulation
272 w/sdurable writes · c=8
41 msp99 latency · c=8near-worst-case write time
2.9Mclient operations checked

FIG. 02 · accretion-db

STORAGE ENGINE · RUST

accretion-db (opens in new tab)

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.

RustLSM treegroup commitcrash consistencyproptest
~29xWAL-bound group commitgain from batching log writes
878 µsfdatasync p50median durable-write flush

FIG. 03 · lodestone

VECTOR SEARCH · RUST

lodestone (opens in new tab)

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.

RustHNSWIVF-PQAVX-512 SIMDRAG retrieval
0.976HNSW recall@10top-10 search accuracy
~31.8Kqueries/sec · 1 core

Eight more: ML, NLP, statistics, and this site

Background

University of Maryland, College Park

B.S./M.S. in Computer Science, M.S. expected May 2027. Departmental Honors (research track)

Aug 2023 – May 2027

On the faculty-mentored research track; minors in Data Science and Philosophy

GPA 3.6/4.0

Coursework: Data Structures & Algorithms, Object-Oriented Programming, Distributed Systems, Operating Systems, Computer Networks, Database Systems, Software Engineering, Machine Learning, Natural Language Processing

Languages

Python · Java · Go · Rust · TypeScript/JavaScript · Scala · SQL · Bash

Distributed & storage

Raft · gRPC/Protobuf · LSM trees · PostgreSQL/SQLite

Backend, web & testing

Flask/FastAPI · Node.js · JUnit 5 · Playwright

ML & NLP

PyTorch · HuggingFace · scikit-learn · LangChain · JAX · Qiskit · DisCoPy · pytket

Cloud & CI

AWS (EC2/VPC, SDK, CDK, CloudWatch) · Docker · Linux · GitHub Actions · CI/CD · Jenkins

Get in touch

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.

Email meivanwang8989@gmail.com