Ayush Ranjan
Notes on machines that learn and systems that don’t.
The archive
·
47 entries
·
2022–present
§ The essays
The archive, in chronological order, with no editorial flourish.
Some of these I’d defend tomorrow morning. A handful I’d quietly retract. I leave them all up because the trail is the point, and because I’d rather argue with a younger version of me than pretend he was already right.
2026
2026 · 05 · 09
Essay · ML systems
Attention without softmax, in plain numbers
2026 · 04 · 22
Field notes · Cloud
Cost-aware autoscaling for inference, or: why p99 lies
2026 · 04 · 03
Essay · Architecture
Your microservice is a distributed monolith
2026 · 03 · 18
Notes · ML
Quantization, with a spreadsheet
2026 · 02 · 29
Case study · ML
When pgvector is enough
2026 · 02 · 11
Short · Architecture
Designing for the second failure
2026 · 01 · 24
Notes
Things I changed my mind about in 2025
2025
2025 · 12 · 14
Essay · ML systems
The shape of a fast eval loop
2025 · 11 · 02
Essay · Cloud
Three failure modes of multi-region inference
2025 · 10 · 21
Essay · Architecture
Boundaries are the design
2025 · 09 · 08
Case study · ML
How a 4-line tokenizer change shifted our eval by 6%
2025 · 08 · 17
Notes
A small grammar for naming services
2025 · 07 · 04
Essay · ML
What KV cache fragmentation cost us this quarter
2025 · 06 · 12
Field notes · Cloud
A cheap, slow, correct way to find a memory leak in a GPU pod
2025 · 05 · 19
Short · Architecture
Two questions before you split a service