Data Science undergraduate at the University of Minnesota (BS, May 2027). I work on LLM serving infrastructure, multi-tenant Kubernetes platforms, and cloud cost attribution — production vLLM endpoints with per-team isolation and token-level billing, a K8s competition platform that served 300+ users, and a first-authored multimodal ML benchmark.
Most of my public work is upstream, in other people's repositories.
12 merged pull requests across 5 projects. Every link below goes to the PR.
vllm-project/vllm — inference engine
- #43362 —
[Bugfix] FusedMoE: coerce shape-(1,) per-tensor scales to 0-D scalar— merged - #43931 —
[Bugfix] V1: clear stale allowed_token_ids mask in InputBatch.condense— in review
vllm-project/aibrix — Kubernetes control plane for LLM serving
- #2404 —
[Bug] Reflect KubernetesJob worker counts and usage on finalize— merged - #2366 —
[Bug] Propagate Redis password to KubernetesJob batch worker— merged - #2331 —
[Bug] Fix per-model metrics cross-talk on multi-model pods— merged - #2450 —
[API] Add explicit StormService deployment mode field— in review
BerriAI/litellm — LLM gateway: routing, guardrails, spend tracking
- #28849 —
fix(customer_endpoints): restrict /customer/daily/activity to admin-only— merged - #28764 —
feat(proxy): allow use_redis_transaction_buffer without redis cache— merged - #28563 —
fix(guardrails): disable_global_guardrails overrides team list— merged - #28260 —
fix(fallbacks): preserve fallback model in SDK fallback responses— merged - #34779 —
fix(complexity_router): escalate a session pin when a turn scores higher— in review
mlflow/skills (Databricks) — agent skills for MLflow
- #18 —
feat: Add Gemini/Google provider support across MLflow skills— merged - #19 —
fix(tracing): flush async trace queue before verifying traces— merged - #20 —
feat(annotate-mlflow-trace): add trace tagging and feedback skill— in review
opencost/opencost (CNCF) — Kubernetes cost allocation
- #3821 —
Add field queryProjectID to BigQueryConfiguration— merged - #3820 —
fix(azure): make ondemand pricing OS-aware for Windows nodes— merged
aws/karpenter-provider-aws — EKS node autoscaling
- #9256 —
fix: set MetadataOptions on the RunInstances dry-run request— in review
WalkBench — first-authored multimodal benchmark for urban walkability. Four cities, 19,624 intersections, built end to end from data pipeline through fine-tuning and evaluation. A multi-task LoRA adapter on SigLIP-SO400M under leave-one-city-out cross-validation; a three-backbone ensemble reaches cross-city Spearman rho = 0.763, and rho = 0.84–0.86 zero-shot on fully held-out Pittsburgh. Workshop paper under review; arXiv preprint August 2026.
agentds-bench — published PyPI package for competition submissions, scoring, and live leaderboards. Built for AgentDS, a Kubernetes platform I solo-built (JupyterHub, 100 isolated pods, per-team storage isolation, automated evaluation) that served 300+ participants in a $10K competition.
Indelible — multi-region tamper-evident audit ledger on Amazon Aurora DSQL. Hash-chained appends over optimistic concurrency control, Ed25519-signed RFC 6962 checkpoints anchored outside the database trust domain, an offline verifier, and an MCP server for AI-agent audit trails. AWS x Vercel hackathon.
GPU MODE (NVIDIA B200) — custom CUDA/Triton Cholesky kernels benchmarked against cuSOLVER; ranked 18/94.
Python · TypeScript · SQL · Bash · C++/CUDA Kubernetes · Docker · Terraform · AWS (VPC, IAM, PrivateLink, S3, EC2) · Azure AKS vLLM · PyTorch · Hugging Face Transformers · FastAPI · PostgreSQL · PostGIS