Animesh Kumar Mishra

Shannon gives you the floor. Every database engine then makes different tradeoffs between compression ratio, read amplification, and CPU cost. Here is the math for four storage architectures.

A note on why I think mathematics is not a language we invented to describe nature, but the structure nature is made of.

A BNPL approval has one hard constraint: the user's thumb is already moving. Here is how the latency budget gets spent, and why the feature fetch — not the model — is the bottleneck.

We ran both in production. The headline numbers are real but the context around them matters more than the benchmarks vendors publish.

On an 8-pod vLLM cluster, p90 TTFT went from 92 seconds to 0.54 seconds with no hardware changes. The only variable was whether the router knew where the KV cache was. This is a cache locality problem