The headline work from my current role. Reduced P95 runtime on a distributed batch grid from 5–8 hours to 10–15 minutes across ~400k positions. Not one change — a sustained program of profiling, cache redesign, removing overweight data paths and eliminating a class of grid node-timeout failures.

Designed and shipped an internal ETL runtime that replaced a third-party tool for a trading-data integration platform spanning 50+ Oracle databases and feeding Salesforce. After a year of hardening in production, it saved ~$900k/year versus the alternative and — importantly — we could debug it ourselves when it misbehaved.

Small internal library that standardized config validation, data validation (with DB lookups), and runtime health checks across a set of Treasury services. Got adopted by 4 teams covering 13 critical processes, and took our weekly investigation count from 1–2 per week to roughly zero. One of those rare tooling projects that paid back its build cost within a quarter.

An Excel VBA module that computes road distances between Russian cities using OSRM (with optional Yandex Geocoder fallback). Built it for a friend's logistics spreadsheet after a conversation that ended with "wait, you can do that in Excel?". Yes, you can. The interface is ugly, the caching is surprisingly good.

An on-and-off evening project: running local LLMs as coding sub-agents (routine file-editing, boilerplate, refactor drudgery) while leaving architectural reasoning to cloud models. Mostly a rabbit hole of hardware trade-offs (RTX 5090 vs. mini-PC form factors), quantization, and CUDA plumbing. I don't have strong conclusions yet, which is part of the fun.