Supply chain and procurement professional with nearly 10 years of experience across purchasing, vendor operations, and enterprise procurement environments. I started automating work in 2016 with Excel VBA macros that cut order upload time from about an hour to ten minutes. At the time, I did not think of it as programming. I thought of it as fixing something that was too slow.
That pattern kept repeating.
In early 2025, a work laptop froze during a large Excel PowerQuery join. Not just slow — full GUI lockup, daily crashes, and no viable path forward inside the tool. That was the point where I moved seriously into Python, tested WSL against native Linux, started profiling the difference, and began working down the stack.
Excel/VBA → Python → Pandas → Polars → Parquet → ramdisk I/O → Linux → syscall profiling.
My focus is data engineering, pipeline performance, automation, and understanding what actually happens underneath high-level abstractions.
The production systems I build at work are confidential, but the engineering patterns are transferable:
A representative production pipeline I built runs daily with no manual intervention, processes hundreds of thousands of records across multiple source systems, and reduced manual task resolution from roughly 30+ minutes to 2–3 minutes per item.
Extracts municipal procurement data from a City of Tempe PowerBI Gov dashboard with no public export path.
The project reverse-engineers the DSR binary protocol used by the dashboard, including bitmask row reconstruction, ValueDict index resolution, and millisecond timestamp conversion. It extracts 9,502 contract records across 928 vendors representing approximately $23B in estimated contract value, then cross-references vendors against USASpending.gov federal award data.
No API access was available. The DSR decode was the only practical programmatic path.
This project demonstrates a core engineering belief: missing infrastructure is a constraint, not a blocker.
I use LeetCode problems as a benchmarking sandbox. When a problem has a one-line solution, I want to know what that line actually costs.
Recent example: From 4 Syscalls to 8,517 — What Doubling a Column Actually Costs Across the Stack
For a simple Pandas column update, I benchmarked multiple Pandas strategies plus NumPy buffer mutation, ASM, C++, Rust, and Polars. I separated warm in-memory execution from cold-start syscall surface using randomized round-robin timing and strace.
Key findings:
*= 2 was slower than = col * 2 in this benchmark..loc[], .apply(lambda), .iloc loops, and .iterrows() all added measurable overhead..to_numpy(copy=False)[:] *= 2 bypassed most Pandas assignment machinery and was the fastest Pandas-backed path.Profile: leetcode.com/u/tUhGYEF4fl
Languages: Python, SQL, Bash, C++ learning, Rust learning Data: Pandas, Polars, NumPy, PyArrow Pipeline: systemd, ramdisk I/O, CDC watermarking, Parquet, stage profiling Automation: Playwright, COM automation Infra: Linux, conda/miniforge, cloud object storage
LinkedIn: patrickryankenneth