writing
Writing
The hard parts of building AI systems in production
Notes from the field. What works, what looks like it works, what actually held up at three a.m. when the on-call alerts started.
Aug 202617 minDNS Was the Internet's First CDNRecursive resolvers are caches, TTLs are cache-control, and the internet's phone book was doing edge delivery fifteen years before Akamai. A first-principles tour of DNS and CDNs — ending with a resolver built from raw sockets and the question of whether Cloudflare has already won.Jul 202624 minI Made Claude Build a Rust Coding Agent With a Team of AI AgentsA persistent coordinator that never writes code, one builder agent per milestone, and a merge gate whose adversarial half caught a real shipping-blocking bug in every single milestone that green CI missed. Here is the whole system — plus the benchmark numbers, methodology, and install once rho shipped.Jun 202615 minCassandra Is a 'Row Store': Truths, Lies, and What the Source Code SaysI cloned the Apache Cassandra repo and read it. What 'partitioned row store' actually means, why a keyspace isn't quite a database, what shape of data Cassandra is actually for, who runs it at what scale, how it compares to ScyllaDB, DynamoDB, HBase and PostgreSQL — plus laptop-scale experiments that make the storage model visible.Jun 202614 minCassandra as a Data Warehouse in 2019: What Worked, What Didn'tIn 2019 at Marax, two engineers stored 10M+ daily click events from a fitness product marketplace in self-hosted Cassandra on Azure, extracted ML features with Spark, and shipped recommendations back over Azure Service Bus. A retrospective on the build-vs-buy call against Cosmos DB, and the two things that hurt most: ops burden and tombstones.Apr 20267 minThe Harness Is the ProductInfrastructure changes consistently outperform model upgrades. TerminalBench proved it. Vercel proved it. Meta paid $2B for it. Here is what actually matters when you are building agents for production.Apr 20266 minMCP vs CLI + Skills: when to use whichThe rule of thumb I use after a year of shipping production agents — and why Andrew Ng's Context Hub ships both a CLI and an MCP server.Mar 20265 minFrom Vibe Coding to Agentic EngineeringVibe coding was fun. Then we had to ship. Here is what actually works when you are building AI agents for production — and why the interesting problems have nothing to do with models.Jan 20265 minHidden Technical Debts in Churn Prediction SystemsResearch poster on improving churn prediction through ranking-based approaches, explainable ML, and segment-aware evaluation - poster presentation at Montreal AI Symposium 2018.