Skip to content
View iwang-1's full-sized avatar

Block or report iwang-1

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
iwang-1/README.md
Ivan Wang, Software Engineer Software systems, backend engineering, and applied ML. SDE intern at AWS, summer 2026. CS B.S./M.S. at UMD, graduating May 2027. Open to Summer 2027 internships and new-grad roles.

AWS SDE Intern, Summer 2026 UMD Computer Science B.S./M.S., May 2027 Open to Summer 2027 internships and new-grad roles

Portfolio: iwang-1.github.io LinkedIn: ivanwang1 Email: ivanwang8989@gmail.com

👋 About

My recent work is centered on three systems written from scratch: a Raft key-value store, an LSM storage engine, and a vector search engine. For each one I wrote the test harness before the thing it was meant to judge.

I'm an SDE intern at AWS on EC2 Networking this summer, finishing a CS B.S./M.S. at UMD on the Departmental Honors research track, and open to Summer 2027 internships and new-grad roles. I'll relocate anywhere in the US.

Every number below comes from a benchmark committed in the repo it describes, so you can re-run it yourself. Where a result is worse than what I compared against, it's published anyway.

Outside of work I'm into mechanical keyboards: you can customize every part, and picking the right one for the job is the fun of it. I also boulder, where a problem can usually be done a lot of different ways, and working out which one costs you the least is my favorite part. And I keep up with where the industry is heading, from new hardware to how governments are handling AI.

🚀 What I built

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.

Batching queued writes into one proposal cut fsyncs from 2.23 to 0.46 per write and took durable throughput at 64 clients from 102 to 1,408 writes/s (p99 764 ms → 60 ms). Production log compaction keeps each node at about 1.3 MB on disk instead of 11.5 MB, and a follower that falls behind catches up through InstallSnapshot.

Go Raft gRPC Porcupine

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. Atomic multi-key WriteBatches are all-or-nothing across another 1,456 crash executions, and index-seeking range scans run 81x faster. Optional background compaction cuts the worst single-writer stall from 760 ms to 208 ms. Building it surfaced three bugs, each fixed with a regression: one in the engine's crash recovery, a GC race, and one in the crash simulator itself. Benchmarked against sled, which wins every matched comparison on this host. Reported rather than omitted.

Rust LSM tree group commit proptest

Vector search engine · Rust

An HNSW proximity graph and an IVF-PQ compressed index over hand-written AVX-512 and ARM NEON 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.

On Apple Silicon, the NEON kernels are 6–12x faster per distance call than scalar code and cut HNSW index build time from 30.3 s to 8.9 s, at identical recall. Both index types save to checksummed files that reload in under 100 ms instead of seconds of rebuilding.

Rust HNSW IVF-PQ AVX-512 NEON

🌍 Upstream & research

4 merged CCD-data-archive

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

open lambeq #259

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

open Big-tech open source

Microsoft mimalloc: found undefined behavior in free(NULL) on ARM with UBSan and reported it with a fix; the maintainer fixed it upstream crediting the PR (9743b25).

Bug fixes under review at Meta RocksDB (a flush crash caused by SetOptions() skipping validation), Google Abseil (a constructor overload bug), and Apple SwiftNIO (a flaky test's thread race). Each one reproduced before fixing.

research 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 ↗

🔧 More projects

Project What it does Result
relaybench · Python An execution lab for tool-using AI agents: inject faults like a lost acknowledgment, compare agent policies on identical runs, replay decisions deterministically. Blind retry double-books in 12/12 seeds, reconcile-first passes 12/12; all 72 runs replay exactly
spillway-join · Rust An external-memory equijoin under a fixed memory budget, with skew recovery. Matches SQLite row for row; adaptive fanout made it 2.6x faster
batchwindow · Python A deadline-aware inference-batching simulator, measured against an exact offline optimum on small traces. Late requests 12 → 4 of 12 vs FIFO, with 30% less device time
leasepost · Python A durable job queue with fenced leases, so a paused worker can never publish a stale result. 256 job histories validated across 8 concurrent workers, including killed processes
minuteledger · Python Event-time windows that absorb late corrections and deletions through a replayable changelog. 40 double-crash trials recover to output identical to a clean run
testtrail · Python Picks the Python tests a Git change could affect and prints the import path behind each choice. Selected 5 of 5 truly failing tests on injected faults in a real project

🧰 What I work in

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

Systems, backend & testing · Raft · gRPC/Protobuf · LSM trees · PostgreSQL/SQLite · Flask/FastAPI · Node.js · JUnit 5 · Playwright

ML, NLP & research · PyTorch · Hugging Face · scikit-learn · LangChain · JAX · Qiskit · DisCoPy · pytket

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

📬 Get in touch

I'm looking for Summer 2027 internships and new-grad software engineering roles, and I'll relocate anywhere in the US. Email is the fastest way to reach me.

Email Ivan Portfolio LinkedIn

Resume on request.

Pinned Loading

  1. lodestone lodestone Public

    A vector search engine in Rust: HNSW and IVF-PQ indexes over AVX-512 and ARM NEON distance kernels, with checksummed index files that reload in under 100 ms.

    Rust 1

  2. parallax-kv parallax-kv Public

    A from-scratch Raft key-value store in Go with deterministic simulation testing, linearizability checking, batched durable writes, and production log compaction.

    Go 1

  3. accretion-db accretion-db Public

    An LSM-tree storage engine in Rust: write-ahead log with group commit, atomic write batches, index-seeking scans, optional background compaction, and a crash-injection harness.

    Rust 1

  4. CCD-data-archive CCD-data-archive Public

    Forked from warnerem/CCD-data-archive

    Web archive for UMD Observatory CCD astronomy data — Python processing pipeline, SQLite + Flask backend, searchable web UI.

    HTML 1

  5. FIRE-QML-WINNERS-QNLP FIRE-QML-WINNERS-QNLP Public

    Quantum NLP sentence classification (UMD FIRE) — pregroup-grammar quantum circuits; enhanced SPSA/ADAM/genetic optimizers beat a default-gain SPSA baseline by ~30 points of train/test accuracy.

    Jupyter Notebook 1 1

  6. iwang-1.github.io iwang-1.github.io Public

    Personal site — new-grad SWE portfolio

    TypeScript 1