// notes

Engineering notes

How I think about AI systems, realtime, and distributed architecture.

Every LLM call creates an ai_runs row with input hash, prompt version FK, token usage, and validation errors for replay and debugging.

Version-agnostic Kafka message schemas, topic naming patterns, delivery semantics, and consumer mapping rules.

Anonymized end-to-end design: client preprocessing, multipart upload gate, Kafka handoff, Go worker transforms, DB URL swap, CDN delivery.

Bounded worker pools in Go media processors: backpressure, retry backoff, result publishing, and graceful drain.

Why probe + policy checks run on completeUpload, retry semantics, object lifecycle, and separation from the transcode worker.

Six separate Node worker entrypoints consume disjoint BullMQ queues so CPU-heavy FFmpeg, GPU transcription, and LLM calls do not starve each other.

Tiered memory, retrieval paths, and the consistency tradeoffs that actually matter.

Fan-out, backpressure, and why your WebSocket layer should be boring.

Why agents are workflows, not loops — and how to keep tool use safe.