Dispatches from the work in progress
The bottleneck of modern vector retrieval. When working with high-dimensional embedding space-whether serving semantic search, recommendation engines, or Retrieval-Augmented Generation (RAG) pipelines - the fundamental challenge is nearest-neighbor search. Exact nearest-neighbor methods require calculating the vector distance (like Cosine or Euclidean) between a query vector and every single item in
The bottleneck in the handoff. For most of this project, converting a Figma screen into Flutter was the slowest, dumbest part of the job. The work wasn't hard, it was just relentlessly manual: open the design, click a frame, read the inspector, note that the gap is 12
Amazon OpenSearch Serverless (AOSS) presents the OpenSearch REST API. You point the standard client at it, sign your requests, and everything you already know appears to apply. That appearance is the whole problem. It is not a managed OpenSearch cluster, it is a different product that accepts a subset of
Building Hardware That Remains Stable After Years of Real-World Operation In early hardware validation, voltage rails usually look perfect. The regulator output is measured at room temperature. The load is controlled. The board passes functional tests. The oscilloscope shows clean waveforms. Then the product spends three years in the field.
The problem Open the Journal tab, wait for it to load, switch to Chat, switch back. For a full beat, Journal says "No journal yet" — then the data you already fetched twenty seconds ago reappears. Do it again. Same thing. Every single time. The device log made it
An API can work perfectly during development and still struggle when the number of users increases. A request that takes a few hundred milliseconds with a small number of users can behave very differently when hundreds of requests arrive at the same time. That made us look beyond functional testing
The symptom Pick a pet photo during onboarding, tap "Use this photo," and land on a scanning screen: "Taking a look at {pet}…" Most of the time it resolves in under a second. Sometimes — almost always on the very first attempt in a session — it just
The hub board inside our feeding station reports air quality, light and motion to the rest of the device over a CAN bus. The firmware queued a frame. The controller accepted it. And nothing — nothing at all — appeared on the wire. There was no error code to chase, no exception,
A notification system that only asks "should this be sent?" will eventually send everything, correctly, all at once and the person holding the phone will switch it off. This post is about the opposite failure, and it's the one users actually complain about: everything arrived, every
Every automated check answers a question about code that exists and paths that run. The defects that survive a mature pipeline live outside both — and they announce themselves the same way a healthy system does.
How clear engineering documentation prevented costly misinterpretation while taking Everbowl from design files to physical hardware. Introduction A CAD model shows what a part looks like. A manufacturing drawing has to communicate what the part actually needs to be. That difference mattered more and more as we moved Everbowl from
What a cross-account DynamoDB migration taught us about the gap between "validated" and "true." A dry-run validates your read path. Only the cutover validates your write path. Everything expensive lives in the gap between the two. We learned this while moving a few hundred DynamoDB tables
"Why is Max scratching more this week?" Somewhere in Max's data, the makings of an answer exist: his bowl logged a food switch on July 5, the new kibble's ingredient list includes chicken meal, his tracker has flagged more nighttime restlessness since July 8,
How working through a simple checklist changed the way I look at web design, development, and testing. While working on a Design-Testing Checklist, I started noticing something I hadn't really thought about before. A web page doesn't become “ready” just because the design is finished. There
The workload, before the database. Every smart bowl and collar in our fleet is a small, relentless firehose. A single device emits roughly twenty distinct event types: camera1 and camera2 frames, thermal and segmented images, audio clips, ocular temperature readings, plus a continuous drizzle of humidity, barometer, imu, proximity, weight,