mitali sinha · india
mitali rust and systems. lately pretty deep in inference.
i learn systems by building them.
currently building marin, a local llm inference engine in rust and cuda.
open to new work
01projects
where does state live, and what survives when something dies?
-
a runtime that keeps an llm chat alive when the worker dies.
proofsurvives real worker sigkills. tested against vllm on a gpu.
still pushing on migration and faster recovery.
how it works
whya crash, a restart or a full gpu usually means the conversation is gone. i wanted a session to have its own id.
insiderust cli + daemon, sqlite wal for history, kv checkpoints on disk. it hibernates idle sessions and hands them off over ssh.
learninga chat is tokens, history and a kv cache that dies with the process.
-
a container runtime with no docker underneath.
proofpid 1 reaps orphans, forwards signals and returns the real exit status.
this rabbit hole got deep fast
how it works
whyi wanted to see what a container actually is.
insideclone() with namespace flags, pivot_root into a private rootfs, a cgroup v2 memory limit.
learninga container is a process with a few lies told to it.
-
git, from scratch, in rust.
proofwrites real git objects (blob, tree, commit), zlib + sha-1 on disk.
95d09f2b10159347eece71399a7e2e907ea3df4fhow it works
whyto learn how git actually stores things.
insideobjects are named by their own contents, so one changed byte changes every hash above it.
learninga commit is just a hash pointing at a hash.
-
a backend for editing the same document at once.
proofbounded queue per client. slow clients get dropped, not waited on.
how it works
whyi wanted to understand how collaborative editors stay consistent.
insidecrdt sync with yrs over websockets and jwt. every room is its own actor and cleans up when the last person leaves.
learningstate lives in the room, not in the socket.
02experience
-
software engineer @ Math Inc
distributed backend and internal infrastructure in rust. storage, networking and developer tools, with reliability as the main constraint. owned features from architecture to deployment.
-
software engineer @ CLR3
real-time data pipelines for on-chain defi analytics. stream processing and rule evaluation on kafka, pyflink, clickhouse and redis, tuned with batching and parallelism so less data moved between services.
03side quests
older stuff, still fun.
- kairoworkflows built from wasm componentsRust (opens in new tab)
- baleriona raw linux tool that watches filesystem eventsRust (opens in new tab)
- saucea typed language that writes like a scriptRust (opens in new tab)
- nexusdependency graphs and cycle detectionRust (opens in new tab)
- kodoa dev activity tracker with a tuiRust (opens in new tab)
- kredentwrite privacy logic in rust, get a mina zkappRust (opens in new tab)
04about
i like the part of software where things break.
failure paths, memory, scheduling, what the machine is really doing.
rust pulled me into low-level systems, systems pulled me into runtimes, runtimes pulled me into inference.
off the clock: manhwa, music on loop, and games when my brain needs a reset. headphones on, “one more round” energy.
the “one more round” part is a lie. it’s never one.