Skip to content
View ksapru's full-sized avatar
11 followers · 2 following

Achievements

Achievement: Pair ExtraordinaireAchievement: Pull Sharkx2Achievement: YOLOAchievement: Quickdraw

Achievements

Achievement: Pair ExtraordinaireAchievement: Pull Sharkx2Achievement: YOLOAchievement: Quickdraw

Block or report ksapru

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.

Content in all repositories owned by your account will be closed.
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
ksapru/README.md

Krish Sapru

Projects

  • h100-inference-control-plane: async FP8 serving on a single H100. Throughput 680 to 3,400 tok/s at 64-way concurrency, p99 latency 4.5s to 1.2s.
  • expense-bench: computer-use agent environment with dual graders. A final-state grader accepts 75.5% of runs vs 56.7% for a path-aware grader, so 1 in 4 apparent successes reward-hacked.
  • per_core_lock_free_event_bus: C++ SPSC ring buffer with cache-line alignment. ~24M msgs/sec at 2.5µs p99.

Open source

  • NVIDIA/cuCollections: allocator lifetime fix in storage deleters, lookup test coverage.
  • NVIDIA/NemoClaw: sandbox readiness races, container secret exposure, CLI test determinism.

Previously SWE II at Intuit (AI Revenue Intelligence) and TPM intern at Google.

krish.sapru.th@dartmouth.edu

Pinned Loading

  1. triton-monitor triton-monitor Public

    Built a GPU inference monitoring and load testing system using NVIDIA Triton, GCP, Prometheus, and Grafana to analyze performance under concurrent workloads.

    Python

  2. h100-inference-control-plane h100-inference-control-plane Public

    H100-targeted async LLM inference control plane for FP8 model serving, observability, SLO analysis, and load testing

    Python

  3. aethel-clean aethel-clean Public

    Off-the-shelf dense retrieval degrades faster than BM25 as corpora grow: a scale-sensitivity study across four general-domain bi-encoders, with an open financial benchmark and a negative result for…

    Python

  4. infra-drift-system infra-drift-system Public

    Forecasting and drift detection suite designed to maintain high-fidelity performance baselines across large-scale infrastructure clusters

    LookML