108 lines
4.7 KiB
Typst
108 lines
4.7 KiB
Typst
// Page setup
|
||
#set page(margin: (x: 0.7in, y: 0.55in))
|
||
#set text(font: "New Computer Modern", size: 10.5pt)
|
||
#set par(justify: true, leading: 0.42em, spacing: 0.42em)
|
||
#show link: set text(fill: rgb("#333333"))
|
||
#set list(indent: 0pt, body-indent: 0.4em, marker: [•], spacing: 5pt)
|
||
|
||
// Section header
|
||
#let section(title) = {
|
||
v(7pt)
|
||
text(size: 11pt, weight: "bold", smallcaps(title))
|
||
v(-2pt)
|
||
line(length: 100%, stroke: 0.4pt)
|
||
v(2pt)
|
||
}
|
||
|
||
// Unified entry: primary header + date, secondary header, then body
|
||
#let entry(primary, date, secondary, body) = {
|
||
grid(
|
||
columns: (1fr, auto),
|
||
text(weight: "bold", primary),
|
||
align(end, text(date)),
|
||
)
|
||
v(-1pt)
|
||
text(style: "italic", secondary)
|
||
linebreak()
|
||
body
|
||
}
|
||
|
||
// === HEADER ===
|
||
#align(center)[
|
||
#text(size: 14pt, weight: "bold")[Sicheng Pan] #h(6pt)
|
||
#text(size: 9.5pt)[
|
||
#link("mailto:sicheng0129@gmail.com")[sicheng0129\@gmail.com] #h(4pt) | #h(4pt)
|
||
+1 (510) 301-0622 #h(4pt) | #h(4pt)
|
||
#link("https://www.linkedin.com/in/sicheng-pan-0129/")[LinkedIn]
|
||
]
|
||
]
|
||
|
||
// === EDUCATION ===
|
||
#section("Education")
|
||
|
||
#entry(
|
||
[University of California, Berkeley], [2023 -- 2024],
|
||
[M.S. in Electrical Engineering and Computer Sciences, GPA: 3.94/4.0],
|
||
[Thesis: _Extensible Rule Language for Query Optimizer_, advised by Alvin Cheung],
|
||
)
|
||
|
||
#v(5pt)
|
||
|
||
#entry(
|
||
[University of California, Berkeley], [2019 -- 2023],
|
||
[B.A. in Computer Science and Statistics, GPA: 4.0/4.0],
|
||
[Highest Distinction in General Scholarship · Honors in Computer Science],
|
||
)
|
||
|
||
// === EXPERIENCE ===
|
||
#section("Experience")
|
||
|
||
#entry(
|
||
[Chroma], [Aug 2024 -- Present],
|
||
[Member of Technical Staff, Data Plane],
|
||
)[
|
||
- Integrated 4-bit RaBitQ quantization into the SPANN index, reducing compaction time from 20 min to 2 min per 1M vectors (1536-dim) with 5× lower memory usage, enabling collections to scale from 5M to 50M with ~40ms query latency and >90% recall\@10.
|
||
- Designed and implemented a hybrid search API supporting composable KNN expressions (e.g., reciprocal rank fusion), sparse vector indexing (BM25, SPLADE) via Block-Max WAND, achieving sub-100ms latency at 1M scale. Shipped end-to-end across engine, API, Python/JS/Rust clients, and docs.
|
||
- Co-led rewrite of the distributed frontend from Python to Rust (tokio/axum), owning the read path (query plan serialization, executors). Throughput increased from 800 to 6,000+ RPS on 16 cores with latency spikes eliminated. Also shipped as the new Rust-based local client.
|
||
- Designed a serializable query plan and pushed query orchestration from the frontend to the query server, reducing network round trips from 3 to 1 per query and eliminating large intermediate data transfers.
|
||
- Implemented instant collection forking with copy-on-write semantics, enabling users to checkpoint datasets and share sample collections without incurring storage copy.
|
||
- Built regex query support via two-stage approach: extracting required literals from regex patterns to narrow candidates via the trigram index, then brute-force matching survivors, achieving sub-100ms latency at 1M scale.
|
||
- Implemented efficient limit/offset pagination, negation filters using roaring bitmaps, CMEK encryption (GCP), and group-by deduplication for chunked-document search results.
|
||
]
|
||
|
||
#v(5pt)
|
||
|
||
#entry(
|
||
[Duolingo], [Summer 2022],
|
||
[Software Engineer Intern, Data Infrastructure and Experimentation Team],
|
||
)[
|
||
- Implemented approximate query pipeline on BigQuery for the analytics dashboard, saving >50% query time at \<1% uncertainty.
|
||
- Implemented caching mechanism for common queries based on historical frequencies (AWS, Jenkins), saving >80% time for analysts.
|
||
]
|
||
|
||
#v(5pt)
|
||
|
||
#entry(
|
||
[R-Polars Project], [Summer 2023],
|
||
[Contributor, Google Summer of Code],
|
||
)[
|
||
- Exported Polars features to R including streaming I/O in Apache Parquet and Arrow formats.
|
||
- Refactored error handling with recoverable errors from Rust and implemented background query pipeline via multi-threading, saving >50% user wait time.
|
||
]
|
||
|
||
// === RESEARCH ===
|
||
#section("Research")
|
||
|
||
#entry(
|
||
[QED: A Powerful Query Equivalence Decider for SQL], [2021 -- 2024],
|
||
[UC Berkeley EECS, advised by Alvin Cheung · Published at VLDB 2024],
|
||
)[
|
||
- Co-developed QED, a SQL query equivalence prover in Rust using a novel formalism (Q-expressions) under bag semantics with a complete checking algorithm for a general query fragment parameterized by first-order theories.
|
||
- Verified 299/444 query rewrite pairs from Apache Calcite and 979/1287 from CockroachDB, more than 2× the coverage of prior state-of-the-art.
|
||
]
|
||
|
||
// === SKILLS ===
|
||
#v(4pt)
|
||
#line(length: 100%, stroke: 0.4pt)
|
||
*Languages:* Rust, Python, SQL, R, Nix \
|
||
*Tools & Infrastructure:* Linux, Git, gRPC/Protobuf, tokio, axum, AWS, GCP, Kubernetes, Docker
|