01 02 03 04 05 06 07 08 09
Kayvane Machine learning engineer writing about systems, ML infrastructure, and new tools.
9 articles · each preview is the system, running
use reduced motion
01 / 09 May 28, 2026 · 7 min
workflow playbooks principles wrapper route · fan out · learn A teardown of pstack by Lauren Tan, and the extensibility move I think will turn skill orchestrations into continual learning loops.
agents agentic-coding tooling claude productivity
02 / 09 May 20, 2026 · 9 min
c1 c2 c3 c4 c5 c6 6 callers claim · entity:42 1 round-trip database 6 callers · 1 fetch Three cache layers in front of an entity-resolution database — in-process LRU, Modal Dict, and the database itself. A distributed lock plus Queue-coordinated notify collapses 50 concurrent containers into one database round-trip per key. With code, animated diagrams, and an interactive simulator.
modal caching system-design python
03 / 09 Jan 17, 2026 · 2 min
svc svc signals observe · decide · evolve How observability tools can lead to a self-evolving product
observability agents system-design
04 / 09 Jan 15, 2026 · 3 min
guard rail plan act verify loops all the way down A field report on agentic workflows: the bitter lesson, guard rails, sub-agent loops, and the evolving role of engineers.
agentic-coding
05 / 09 Jan 13, 2026 · 4 min
unread, unpaid raw repo context tl;dr 14,200 tok 3,400 tok 380 tok same context · 2.6% of the tokens
A tutorial for installing llm-tldr with uv and wiring Claude hooks for instant, structured context, plus real numbers from this repo.
llm tooling claude productivity
06 / 09 Jan 12, 2026 · 7 min
seq a seq b seq c shared prefix paged · no fragmentation
A practical tour of vLLM's LLMEngine, scheduler, and paged KV cache, plus why paging and radix trees drive throughput.
llm vllm gpu
07 / 09 Jan 08, 2026 · 10 min
0ms 200ms 400ms 600ms 800ms handler db.query modal.fn · 340ms sampled p95 340ms · traced end to end
A practical walkthrough for wiring Datadog APM into Modal, with tracing pitfalls and sampling tradeoffs.
observability modal datadog python
08 / 09 Jan 02, 2026 · 1 min
farkle · one throw +100 +50 +100 +100 +50 +100 +350 +450 450 · bank or roll?
We speed-ran a playable Farkle web game in about an hour using GPT-5.2-Codex for the core build, GPT Image Gen for character art, and Suno for a Game Boy-style theme.
llm game-dev gpt suno image-gen
09 / 09 May 13, 2024 · 9 min
regex → fsm { { " " a-z a-z " " : : } } masked tokens output { { " " name name " " : : "ada"} "ada"} valid json · guaranteed A walkthrough of Outlines and finite-state machines for constrained LLM generation, with regex and Pydantic examples.
llm outlines structured-generation pydantic
01 Skill orchestrations, and what I like about pstack May 28, 2026 02 Claiming, not caching — singleflight on Modal May 20, 2026 03 Event Driven Evolutionary Software Jan 17, 2026 04 On building with agents Jan 15, 2026 05 LLM-TLDR + Claude Hooks: Fast Context Without the Token Tax Jan 13, 2026 06 How vLLM Works Jan 12, 2026 07 Setting Up Datadog APM in Modal Jan 08, 2026 08 Building a mini game with GPT-5.2-Codex, Suno and GPT Image Gen Jan 02, 2026 09 Guided Generation with Outlines May 13, 2024 © 2026 Kayvane Shakerifar · every article is a system; each preview is that system running · built with TanStack Start & Motion · set in Archivo & Fragment Mono