Skip to content
View riskywindow's full-sized avatar

Achievements

Achievement: Pull SharkAchievement: QuickdrawAchievement: YOLO

Achievements

Achievement: Pull SharkAchievement: QuickdrawAchievement: YOLO

Block or report riskywindow

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
riskywindow/README.md

Rishi Vinodkumar — ML systems, inference, compilers, and distributed systems

CS + Math at UT Austin. I build performance-critical ML infrastructure across inference serving, compilers, distributed systems, and online learning.

LinkedIn  ·  Email

Selected systems

01 · SLOForge

Inference deployment compiler + adaptive runtime

Profiles deployment candidates, searches constrained configurations, adapts online, and validates behavior through replay, fault injection, and machine-readable evidence.

profiling optimization runtime adaptation

02 · FissionSpec

Outcome-decoupled speculative LLM serving

Per-request scheduling, fission and re-coalescing, an exact sampling oracle, transactional KV state, and reproducible CPU experiments.

continuous batching KV state Rust

03 · KEA

End-to-end int8 inference accelerator stack

A custom ISA, cycle-approximate simulator, MLIR compiler, post-training quantization frontend, and reproducible model execution.

MLIR computer architecture C++

04 · Epochless

Correctness infrastructure for online-RL serving

Failure-atomic publication, causal attestation, compact proofs, crash recovery, and formal TLA+ models for continuously updated serving.

formal methods fault tolerance online RL

05 · Switchyard

Hybrid inference control plane

Typed routing, admission, fallback, canary, and evidence semantics across local and remote model backends, with deterministic replay.

model serving control plane replay

06 · Bifrost

Correctness-first KV-state transport

Content-addressed objects, Python/Rust validation, LMCache remote storage, fault injection, and replayable benchmarks.

KV cache distributed storage Rust

Working set

Python, Rust, C++, Go, PyTorch, Docker, Kubernetes, and Linux

LLM serving · ML systems · compilers · distributed systems · performance engineering

Currently working across production infrastructure and ML systems research.

Pinned Loading

  1. Nexus Nexus Public

    C++23 order-matching engine with price-time priority, lock-free SPSC pipelines, zero-allocation hot paths, and io_uring asynchronous UDP networking.

    C++ 1

  2. Bifrost Bifrost Public

    Correctness-first KV-state transport and storage for LLM serving with content-addressed objects, cross-language validation, fault injection, and replayable benchmarks.

    Python

  3. Epochless Epochless Public

    Correctness-oriented lifecycle infrastructure for online-RL serving with failure-atomic publication, causal attestation, compact proofs, and crash recovery.

    Python

  4. FissionSpec FissionSpec Public

    Outcome-decoupled continuous batching for speculative LLM serving with per-request scheduling, transactional KV state, and reproducible CPU experiments.

    Python

  5. KEA KEA Public

    End-to-end int8 inference accelerator research stack with a custom ISA, cycle-approximate simulator, MLIR compiler, PTQ frontend, and reproducible model execution.

    C++

  6. slo-forge slo-forge Public

    SLO-driven compiler and adaptive runtime for LLM inference deployments with profiling, constrained optimization, replay, fault injection, and evidence-backed validation.

    Python