Engineering notes · measured, not claimed

Systems writing, with the arithmetic left in

Twenty-one long-form write-ups from work I have shipped: sizing an ML platform from public traffic numbers, logical clocks that survive concurrent merges, retrieval pipelines I ran end to end, and the interview rounds that tested whether I could teach any of it. Each page keeps its real numbers, its failure modes, and the part I would do differently.

The four tracks

01 · Trip @ Scale

Sizing a consumer AI product

Traffic-to-capacity math, customer segmentation, tag clustering, prompt-cache economics and the RAG architecture that ties them together — all anchored on Tripadvisor's public numbers.

7 articles

02 · Senior SDE

Distributed primitives, in the weeds

Hybrid and vector clocks for CRDT merges, what a Postgres sequential scan actually costs, and shipping rate limiting to Redis as a Lua script.

4 articles

03 · Forward Deployment Engineer

Shipping retrieval for customers

A runnable stdlib-only RAG pipeline, vector-database capacity planning, data-size estimation, the FDE role itself, and a resume built for it.

9 articles

04 · Interview Experience

What the rounds actually asked

A transcript-level account of applied-AI interview loops, the questions that separated candidates, and how I would run the same round.

1 article

How to read these

Every article is first-person and self-contained: I state the constraint I was working under, do the arithmetic in the open, then list the failure modes I design against. Where I ran code, I pasted its real output — including the runs that looked bad. Where a number is an estimate derived from public data rather than a disclosed figure, I say so at the point of use.

Start with User Stats if you want the estimation discipline, Complete RAG Tutorial if you want something to run in the next five minutes, or Vector Clocks for CRDT if you want the deepest rabbit hole.