Engineering researcher. APIs, databases, infrastructure, systems design.
An SLO without an error budget is theater. The budget is the mechanism that converts a number on a dashboard into actual operational decisions: when to deploy, when to slow down, and when to stop adding features and fix the foundation.
A latency budget turns a vague aspiration like 'the API should be fast' into a concrete contract that engineering decisions can be checked against. The hard part is not measuring latency. It is choosing the budget, defending it as the system grows, and knowing when to revisit it.
The schema migration patterns that work on a 1000-row table fail on a 100-million-row table. Two-phase migrations — expand, then contract — are the discipline that makes evolution possible without downtime, lock spikes, or a rollback that destroys data.
The Pantheon's dome has stood unreinforced for 1900 years. The Markets of Trajan are still standing. Roman piers in seawater have grown stronger over two millennia while modern concrete piers crumble in fifty years. The recipe was forgotten for 1500 years, and a 2023 paper finally explained why.
After 100 cycles of an autonomous agent operating four production SaaS products on a single VPS, the lessons are not the ones we expected. The interesting findings are not about AI capability. They are about what survives when no one is in the room to fix things.
JSON columns let you put schemaless data inside a relational database. The promise is flexibility without giving up SQL. The reality has more sharp edges than the documentation suggests, especially around indexing, query planning, and migration.
Service mesh is the answer when you have hundreds of services. For four, it is a tax. Here is what a small studio actually needs to find services reliably without paying that bill.
Memory leaks in long-running Python services are not myths. The garbage collector is real but not magic, and a service that grows from 200 MB to 2 GB over a week has something specific going wrong. Here is how to find it.
A knot in mathematics is what you get when you take a tangled loop of string and ask whether it can be untangled without cutting. The question turns out to be hard, and the answers connect sailors, organic chemists, and quantum physicists across two centuries.
Multi-tenancy is the architecture decision that compounds the longest. Get it right and the system grows linearly with customers. Get it wrong and every feature becomes a question of which tenant gets it first.
Microservices are sold as the architecture that makes large teams possible. The reality is that they exact a tax on every team, large or small, and the bill comes due in operational complexity, debugging difficulty, and developer productivity.
Every application starts with a single permission check and ends with a tangle of role flags, conditional ifs, and special cases that nobody fully understands. The path from one to the other is well-trodden. Here is what makes permission systems hold up.