Series
14 posts
10 posts
When a model answers over private knowledge, the answer must never leak what it used. How to scrub IRIs, UUIDs, and query fragments across streaming chunk boundaries, and how to project provenance as an allowlist.
Dense embeddings find paraphrases; full-text finds exact terms. How to run both legs concurrently and fuse their rankings with RRF — a simple formula that consistently beats either leg alone.
Background jobs that run for minutes, pause for human answers, and survive restarts need more than a status column. A design for pausable, resumable, auditable background work: explicit states, a question channel, generation gates, and append-only logs.
What if a language model compiles a procedure once and never runs again? A skill compiler with hard structural gates, a zero-model interpreter that walks the graph in plain code, and a trust ladder that must be earned run by run.
How to make a language model reliably drive a structured UI: force a tool call every turn, treat the final answer as a schema, feed validation errors back for self-correction, and bound the whole thing by construction.
How do you compare LLMs on generating structured queries? Syntax similarity is worthless. What works: executing generated queries against a live triplestore and comparing result sets — plus a spread rule most benchmarks ignore.
Third-party webhooks expire, get orphaned by restarts, and miss events. A reconcile-loop design that converges desired vs. live state, plus deduplication by join URL and latches that remember deliberate exits.
How does an assistant remember you across sessions? A speaker-attributed extraction pipeline with dedup and rate limits, three isolation scopes, and a retrieval hierarchy that ranks topics, decays by recency, and never truncates silently.
Should a survey be "completed" or "expired"? Store the flag and it lies between updates; compute it on read and it is always true. A design for lifecycle derivation, assignee coverage, and the short-circuits that keep it fast.
Voice agents feel synchronous but live on a mess of tool calls, reconnects, and long-running jobs. A background job manager with dedup, bounded retries, a dead-letter queue, and a sideband control channel — plus a discovery system for assumptions you cannot verify.
4 posts
Two small algorithms that make a notification service boring in the best way: cursor pagination that cannot skip rows, and exactly-once-flavored dispatch built on a single SQL statement.
How to run a distributed scheduler on a single table: advisory-lock leader election, SKIP LOCKED work claiming, semaphore-style concurrency caps, and heartbeat reaping.
iCalendar RRULE looks like a solved problem until INTERVAL=0 hangs your sweep, BYHOUR means the wrong hour across timezones, and legacy rows encode their whole schedule in one string. A field guide to doing it safely.
The small but sharp security machinery behind an API endpoint: in-memory token buckets, peppered SHA-256 with constant-time comparison, and deterministic session namespacing.