early career #
- 2019 Dask on Azure ML Led OSS contribution enabling distributed data processing and ML training.
- Mar 2020 Azure ML v2 Led data experiences across the platform rewrite: CLI, web API, Python SDK, and more.
- 2021 DX repair loop Overhauled docs/examples and turned enterprise feedback into product fixes.
- 2022 MLOps features Owned Spark integration and shipped event-driven automations, Git integration, batch inferencing support, and more.
Micro ECG
Reverse-engineered the proprietary storage format to recover the signal.
00004D 45 43 47 01 00 80 3F 7A 12 C4 09 5E 00 11 2D
00107F 00 82 00 91 00 A8 00 C4 00 F1 00 2E 01 68 01
00209B 01 B0 01 A4 01 77 01 25 01 C8 00 88 00 6B 00
003061 00 69 00 7C 00 8E 00 94 00 87 00 70 00 5D 00
004055 00 5C 00 77 00 B5 00 44 01 E8 01 6A 02 01 03
00507D 02 B2 01 EC 00 87 00 63 00 5F 00 67 00 74 00
006080 00 8B 00 90 00 87 00 73 00 60 00 57 00 5B 00
Hospital alarm mesh
IRB-approved study at Shands: Raspberry Pis with microphones detecting non-critical alarms and routing alerts to the nurse station.
non-critical alarm
room 412 · detected
trend stable
route to station
Azure ML data drift
Observed production data moving away from training data and turned that into an actionable cloud ML monitoring surface.
feature distribution moved past threshold
GitHub Copilot beta
Early AI-coding adopter: testing the workflow when it still felt experimental, not inevitable.
def route_signal(graph, source):
frontier = [source]
while frontier:
node = frontier.pop()
yield from model.suggest_next(node)Full-stack SoC tooling
Two browser frontends over simulation backends, plus low-cost instrumentation when direct printf made slow SoC simulations even slower.
HTML/CSS/JS
Perl
Python
latency
power
area
HTML/CSS/JS
Python
C
e HDL
format string
debug buffer
end-of-sim readout
Avoided per-cycle printf overhead in simulations already orders of magnitude slower than software.