APPLICAT ION
FOR
GRADUAT E
ADMISSION
·
FALL
2026
Personal
Statement
Master
of
Science
in
Computer
Science
—
Machine
Learning
Systems
Maya
Chen
·
University
of
Washington
The
summer
I
turned
nineteen
,
I
spent
three
weeks
debugging
a
memory
leak
in
a
natural
-
language
parser
I
had
built
for
a
volunteer
project
mapping
humanitarian
crises
.
I
was
working
with
a
dataset
of
400,000
unstructured
news
reports
on
a
single
AWS
t
2.
micro
instance
.
The
model
worked
beautifully
on
1,000
rows
and
crashed
predictably
on
10,000.
That
tension
—
between
elegant
algorithms
and
the
stubborn
physics
of
real
-
world
computation
—
has
driven
every
decision
I
have
made
since
.
I
am
applying
to
the
M
.
S
.
in
Computer
Science
program
at
the
University
of
Washington
because
I
want
to
move
beyond
applying
machine
learning
models
and
begin
designing
the
systems
that
make
responsible
,
large
-
scale
ML
possible
.
I
graduated
from
UC
Berkeley
in
May
2021
with
a
B
.
S
.
in
Electrical
Engineering
and
Computer
Sciences
and
a
concentration
in
Machine
Learning
,
completing
the
degree
with
a
GPA
of
3.91.
My
coursework
in
Computer
Architecture
,
Distributed
Systems
,
and
Statistical
Learning
Theory
gave
me
a
dual
fluency
in
hardware
-
aware
optimization
and
probabilistic
reasoning
.
In
my
senior
thesis
,
advised
by
Professor
Joseph
Gonzalez
,
I
explored
quantization
-
aware
training
strategies
for
edge
-
deployed
computer
-
vision
models
.
By
combining
structured
pruning
with
mixed
-
precision
scheduling
,
I
reduced
model
size
by
62%
while
preserving
top
-1
accuracy
within
1.2%
of
the
full
-
precision
baseline
on
the
COCO
object
-
detection
task
.
The
work
was
accepted
as
a
poster
at
the
TinyML
Research
Symposium
and
taught
me
that
performance
is
not
an
abstract
metric
but
a
negotiated
settlement
between
mathematics
,
silicon
,
and
energy
budgets
.
After
graduating
,
I
joined
Stripe
as
a
Machine
Learning
Engineer
on
the
Fraud
Intelligence
team
.
Over
three
years
,
I
progressed
from
building
batch
inference
pipelines
to
leading
the
redesign
of
our
real
-
time
transaction
-
scoring
architecture
.
My
most
consequential
project
involved
migrating
our
core
gradient
-
boosted
fraud
model
from
a
daily
batch
job
to
an
online
inference
service
backed
by
DynamoDB
and
Triton
Inference
Server
.
I
introduced
feature
-
store
-
backed
consistency
checks
and
model
-
parallel
serving
that
reduced
p
99
latency
from
340
milliseconds
to
52
milliseconds
and
improved
detection
recall
by
3.4
percentage
points
—
translating
to
an
estimated
$
14
million
in
annual
prevented
fraud
losses
.
This
work
required
me
to
operate
simultaneously
at
the
level
of
Bayesian
calibration
curves
and
Linux
kernel
TCP
tuning
,
a
duality
that
convinced
me
my
long
-
term
contribution
lies
at
the
intersection
of
systems
and
ML
research
.
Outside
of
Stripe
,
I
have
maintained
a
collaboration
with
the
Berkeley
RISELab
,
contributing
to
the
vLLM
project
since
late
2022.
I
implemented
a
paging
-
aware
KV
-
cache
manager
that
improved
throughput
for
long
-
context
LLM
serving
by
28%
under
heavy
load
.
The
patch
was
merged
in
release
v
0.3.0
and
is
now
deployed
in
several
production
environments
.
Through
this
open
-
source
work
,
I
1 / 2
have
learned
that
the
most
impactful
research
is
the
kind
that
ships
:
peer
-
reviewed
,
benchmarked
against
real
workloads
,
and
documented
well
enough
for
downstream
engineers
to
adopt
without
pain
.
While
my
professional
work
has
given
me
deep
operational
expertise
,
I
want
to
step
back
and
investigate
the
foundational
questions
I
could
only
glimpse
in
production
.
How
should
we
design
memory
hierarchies
for
trillion
-
parameter
models
when
DRAM
bandwidth
is
the
binding
constraint
?
What
scheduling
guarantees
can
we
provide
for
ML
training
workloads
running
on
preemptible
cloud
spot
instances
without
sacrificing
convergence
bounds
?
These
questions
sit
at
the
boundary
of
computer
systems
and
statistical
theory
,
and
answering
them
rigorously
requires
the
concentrated
time
,
mentorship
,
and
intellectual
community
that
only
a
top
-
tier
graduate
program
can
provide
.
The
University
of
Washington
is
my
first
choice
because
its
systems
-
ML
ecosystem
is
uniquely
cohesive
.
I
am
eager
to
work
with
Professor
Luis
Ceze
on
hardware
-
software
co
-
design
for
neural
network
acceleration
,
and
I
have
followed
Professor
Magdalena
Balazinska
’
s
work
on
provenance
and
metadata
management
in
ML
pipelines
since
reading
her
VLDB
2022
keynote
.
The
Systems
and
Machine
Learning
Lab
’
s
emphasis
on
reproducible
evaluation
—
publishing
not
just
papers
but
fully
open
artifact
sets
—
mirrors
my
own
belief
that
research
integrity
is
inseparable
from
engineering
transparency
.
Courses
such
as
CSE
599
M
:
Distributed
Systems
for
ML
and
CSE
590
D
:
Seminar
on
ML
Systems
would
directly
deepen
the
theoretical
frameworks
I
need
to
complement
my
industry
experience
.
After
completing
the
M
.
S
.,
I
intend
to
pursue
a
Ph
.
D
.
focused
on
sustainable
and
efficient
ML
infrastructure
,
with
the
goal
of
either
joining
a
faculty
position
or
leading
a
research
lab
in
industry
.
Long
-
term
,
I
want
to
contribute
to
open
-
source
standards
that
make
efficient
ML
the
default
rather
than
the
exception
—
similar
to
how
the
LLVM
project
unified
compiler
backends
,
but
for
model
serving
and
training
runtimes
.
Whether
in
academia
or
industry
,
I
hope
to
mentor
engineers
from
non
-
traditional
backgrounds
and
to
advocate
for
research
practices
that
prioritize
reproducibility
and
environmental
impact
alongside
raw
benchmark
wins
.
I
am
applying
to
graduate
school
not
because
I
am
dissatisfied
with
my
career
,
but
because
I
have
reached
the
edge
of
what
I
can
teach
myself
between
pager
-
duty
rotations
.
The
problems
I
want
to
solve
next
—
inference
at
the
planetary
scale
,
energy
-
aware
training
schedules
,
robust
distributed
learning
under
heterogeneity
—
require
the
structure
,
rigor
,
and
collaborative
density
of
a
program
like
the
University
of
Washington
’
s
.
I
am
ready
to
make
that
commitment
,
and
I
would
be
grateful
for
the
opportunity
to
contribute
to
your
community
.
2 / 2