I batched three five-second video futures from one frame on a B200, built a choice-and-continue interface, and found that fast generation and controllable outcomes are very different things.

Quantizing a video model's linear layers to 4-bit cut a 1080p generation by 39 seconds — and did nothing at all on the next two models I tried it on. The gap is the interesting part.

How I rebuilt a production log triage system — binary search over a sorted index, a channel-based ingestion pool, a write-ahead log for crash recovery, and what I'd do differently at scale.

How I built an end-to-end log anomaly detection pipeline that combines classical ML with LLM summarization — and what I learned about evaluating LLM output in production-adjacent systems.