Coding

Last quarter I had to ship a support triage bot for a new product line. We had maybe 30 labeled tickets, no data pipeline, and a hard deadline because the support team was drowning. I could have built a tiny classifier and asked for more labels, but that would not land in time. Instead I […]

I still remember the first time a model generated an image that looked almost real, and I had that “wait, did a human make this?” moment. That feeling is why I keep returning to GANs. They’re the rare model family that doesn’t just classify or predict; they create. When I work with GANs, I’m not

Last month I opened a repo to hotfix a billing bug and ran git stash list. I expected one or two entries. I saw nineteen. Each one had a vague message like "WIP on feature/search-api". That list is a stack of sticky notes, and once it gets tall you stop trusting what is inside. I’ve

The first time I tried to model breathing for a medical device prototype, I treated lung pressures like a single sensor value. The data immediately looked wrong: the pressure I expected to be negative was positive, and the value that should have oscillated was flat. That mismatch forced me to learn the two pressures that

Last quarter I reviewed a deployment that failed on a Linux host with /bin/bash^M: bad interpreter. The script was fine, the path existed, yet the shell stopped at the first line. The culprit was invisible: the file came from Windows and every line ended with CRLF. In a codebase with thousands of files, that tiny

Last year I worked with a product analytics team that needed to split traffic across five recommendation strategies. The numbers had to stay between 0 and 1, and the total had to be exactly 1. We tried normal and logit‑normal models, but we kept tripping over the sum‑to‑one constraint and awkward parameter interpretation. The Dirichlet

Every time a video call stutters or a game jumps, I see teams blame cloud latency, yet the bottleneck is often a living-room router or a crowded office access point. As a senior engineer, I treat Wi-Fi like any other shared system: understand the protocol, measure it, then decide whether to upgrade. You’re probably sitting

Last quarter I helped a team debug a scheduling engine that kept producing impossible timetables. The engine solved a set of linear constraints, but a tiny sign error in the elimination step flipped two rows and the final answer looked valid while being wrong. The fastest way to see the bug was to write the

Last weekend I queued up a Korean thriller after a long day, and the background hum in my living room turned half the dialogue into mush. Subtitles saved the night, but only because VLC could locate the right file in seconds. When captions fail, it is rarely VLC itself; it is usually a mismatch between

Last quarter I watched a small startup lose a week arguing about which AI assistant to standardize on. Meanwhile their backlog of tests, docs, and customer replies stayed frozen. The lesson for me was simple: ChatGPT and Google Bard are not a single choice, they are two different toolchains with different strengths. Bard has since

Procurement used to feel like a stack of email threads, disconnected spreadsheets, and approval bottlenecks. When I walk into a large enterprise today, I still see teams with strong engineering habits but weak purchasing plumbing: a requisition can be tracked in Jira, while the actual purchase order lives in an inbox and the invoice sits

Last year I joined a migration where a retail team moved from VM-based deploys to containers. The first week looked fine, then production buckled because the on‑prem cluster had no integrated registry and the cloud cluster applied a stricter security rule that blocked the image from running as root. The team didn’t fail at writing