ENGINEERING DOMAIN
Data & Storage
Databases, storage systems, indexing and data infrastructure.
Databases, storage systems, indexing and data infrastructure.
5 projects · 6 pieces of writing
Content StudioSelf-hosted content operations platform for drafting, review, approval, scheduling, workspace isolation, audit history and durable AI generation workflows.
Disaster Recovery & Restore VerificationEncrypted backup and recovery system spanning local rsync, off-site object storage, checksums and disposable restore verification rather than backup-success alone.
Infrastructure Operations PlatformCentral service, deployment, policy, storage and execution control plane with RBAC, audit events, read-only command views and guarded host execution.
Social PublisherDedicated publishing engine with scheduler, queue worker, PostgreSQL/Redis state, reproducible source pinning and restore-tested backups; external publishing remains deliberately disarmed until credentials are configured.
Technical Publishing & Portfolio StackFirst-party Flask/PostgreSQL publishing system powering the technical archive, case studies, admin workflow and public portfolio, with a parallel Tor v3 publishing path.
Cloudflare KV, Durable Objects, Queues, and D1 Fail Differently Because Their Consistency Models DifferA lost counter, duplicate job, stale read, or runaway write bill is usually not 'a Cloudflare bug.' Each storage primitive makes different consistency and delivery promises, so the same application pattern fails differently on each one.
A Distributed Lock Without Fencing Can Let a Dead Worker Come Back and Corrupt StateA lease can expire while its old owner is paused rather than dead. If that process later resumes, the lock service may have moved on—but the storage system can still accept stale writes unless it has a fencing mechanism.
pgvector HNSW Can Return Fast Answers That Are Operationally WrongApproximate nearest-neighbor search is allowed to trade recall for speed. Add tenant or metadata filters after the ANN scan and a fast query can return too few candidates without being a database error.
PostgreSQL VACUUM Reclaims Space Without Giving the File Back to the FilesystemStandard VACUUM makes dead-tuple space reusable inside a table. It usually does not shrink the table file, which is why a successful vacuum and unchanged disk usage can both be correct.
Why Observability Workloads Keep Ending Up in ClickHouseLogs and traces are wide, append-heavy events that are usually filtered by time and a few fields, then aggregated. That workload shape explains the attraction to a columnar analytical database better than any benchmark headline.
ZFS Scrub, Resilver, Snapshot, and Replication Solve Different Failure ModesA scrub detects latent corruption, a resilver reconstructs missing redundancy, a snapshot preserves an earlier dataset state, and replication puts that state somewhere else. Calling all four 'backup' hides the failures each one cannot solve.