Team
TL (Tech Lead, not people-manager) of Codebase Awareness team for Gemini-in-Android-Studio, 4 engineers
Location: Kirkland, WA, US
Contact
E-mail: duke.acacia at gmail.com
GitHub: https://github.com/jaebaek
Compiler and systems engineer turned coding-agent tech lead. Currently lead the search tooling
for a shipping coding agent at Google; independently designed and built an end-to-end
verifiable-reward (RLVR) training environment — oracle, task corpus, and RFT loop —
on a single GPU.
Work Experience
Google, TL of Codebase Awareness team for Gemini-in-Android-Studio (June 2025 ~ present)
-
2026 Q2: led a token-efficiency initiative across the agent's tool-call surface. Reduced
average tool-call response size by 30%+ (measured on real user traffic)
and, combined with a reduction in agent step count, cut end-to-end token usage by
~70% on eval — without a capability regression.
-
Since 2026 Q1, TL for all search tooling used by the Gemini-in-Android-Studio (GiAS)
coding agent, as Tech Lead (Google separates the TL and people-manager roles; this is TL
only) for a team of 4 engineers.
-
Built the evaluation suites the team hill-climbs on, and drove retrieval quality against
them.
-
Own the Android Knowledge Base (AKB) retrieval tools for GiAS — first-party API and
Gradle build knowledge surfaced to the agent — in collaboration with the Android
DevRel team.
Google, Kotlin compiler/IDE team (Jun 2022 ~ May 2025)
-
Supported Android Studio features including Jetpack Compose editing, LiveEdit, and the Compose compiler for the new Kotlin (K2) compiler.
-
Contributed to the IntelliJ and Kotlin compiler open-source projects — IDE features and Kotlin compiler bug fixes, in particular those triggered by the Compose compiler plugin.
Google, Vulkan GPU shader compiler team (Mar 2018 ~ May 2022)
-
Vulkan is a next-generation graphics and compute API for GPU applications. Worked primarily on compilers for IR generation and IR optimization.
-
DXC: HLSL to SPIR-V compiler based on Clang (136 PRs, TL Feb/2021 - May/2022).
-
TL compiler part of Vulkan shader debugger: debug info generation from front-end compiler, preservation through all optimization passes.
-
spirv-opt: SPIR-V optimization passes (80 spirv-opt/spirv-val PRs).
-
Others: SPIR-V extension specs e.g., OpenCL.DebugInfo.100, and shader testing tools (62 Amber commits).
Google, Chrome for Android (Apr 2017 ~ Feb 2018)
-
Engineered multiple layers from Blink to UI to merge device
and page zoom scale factors (almost 40 commits).
See
this bug.
Independent Research
Automatically-derived verification oracles for RLVR-based coding-agent training (2026)
-
Designed and built a method that derives a behavioral-equivalence oracle directly from an
existing (legacy) implementation, rather than requiring a hand-written test or oracle per
task — the standard approach in comparable benchmarks, and one that does not scale
past a few hundred hand-authored tasks.
-
Built the environment end to end on a single GPU: runtime capture, a cross-toolkit
normalized tree representation, a tolerance-calibrated multi-channel comparator covering
both static structure and live interaction traces, a hermetic containerized grader, and
the rollout and training loop on top of it.
-
Ran full SFT/RFT rounds against the generated corpus, root-causing pipeline and comparator
defects by reading actual failing trajectories rather than guessing.
-
Measured a concrete reward-fidelity failure mode: decomposing the work into
one-unit-at-a-time subtasks let a model satisfy the grading rule while leaving reactive
state updates disconnected — invisible to a static-structure check alone.
Root-caused it, quantified it by re-grading every previously collected attempt with
interaction-level checking enabled, and ran a follow-up experiment testing whether a cheap
prompt fix resolves it without further training.
-
Validated the oracle in both directions — not only the attempts it passed, but every
attempt it rejected, audited independently against rendered output to confirm each
rejection was a real defect and not a false negative.
Publications
CHANCEL: Efficient Multi-client Isolation Under Adversarial Programs
[ slides |
video]
Adil Ahmad, Juhee Kim, Jaebaek Seo, Insik Shin, Pedro Fonseca, Byoungyoung Lee,
Proceedings of the 2021 Network and Distributed System Security Symposium
(NDSS '21), San Diego, CA, US, February 2021.
(Acceptance ratio: 15.2%)
- Designed the two-way sandbox mechanism for multi-threaded enclave programs
- Implemented LLVM X86_64 backend for instrumentation
- Implemented dynamic ELF loader from scratch
SGX-Shield: Enabling Address Space Layout Randomization for SGX Programs
[
slide-pdf |
slide-ppt]
Jaebaek Seo, Byoungyoung Lee, Seongmin Kim, Ming-Wei Shih, Insik Shin, Dongsu Han,
Taesoo Kim,
Proceedings of the 2017 Network and Distributed System Security Symposium
(NDSS '17), San Diego, CA, US, February 2017.
[ BibTex ]
(Acceptance ratio: 68/423=16.1%)
- Designed and implemented fine-grained ASLR (32-byte granularity)
and software-based Data Execution Prevention (DEP) in SGX
- Implemented dynamic loader and static linker from scratch
- Engineered LLVM backend for X86_64
-
Source code in github:
SGX-Shield
FLEXDROID: Enforcing In-App Privilege Separation in Android
[
slide-pdf |
slide-ppt]
Jaebaek Seo, Daehyeok Kim, Donghyun Cho, Taesoo Kim, Insik Shin,
Proceedings of the 2016 Network and Distributed System Security Symposium
(NDSS '16), San Diego, CA, US, February 2016.
[ BibTex ]
(Acceptance ratio: 60/389=15.4%)
-
Designed and implemented an extension of the Android permission system
-
Protected threads in the Dalvik VM from Java Native Interface (JNI) calls
in malicious third-party libraries, using in-process memory
isolation based on ARM domains
-
Implemented inter-process communication using a character device driver
-
Engineered the Dalvik VM, Android's linker, and kernel page tables to
support ARM domains
-
Source code in github:
FLEXDROID
Education
Ph.D. from Computer Science, KAIST Mar 2013 - Aug 2017
M.S. from Computer Science, KAIST Mar 2011 - Feb 2013
B.S. from Computer Science, KAIST Mar 2006 - Aug 2010
(KAIST stands for Korea Advanced Institute of Science and Technology.)
Other Experiences
-
Presented SGX-Shield at Intel's research lab, Hillsboro, Oregon (Feb 2017).
-
Visiting researcher, Systems Software & Security Lab, Georgia Tech
(Mar – Apr 2016) — SGX-Shield, with prof.
Taesoo Kim and prof.
Byoungyoung Lee.
-
Research intern, Microsoft Research Asia (2011–2012), Mobile and Sensor
Systems group — cloud gaming. Software engineering intern,
Google Korea (2010), Blogger team.