AA

6×

Hackathon Winner

CMU 3× / Harvard / Berkeley / Claude

Engineer / Researcher / Builder

Aarush
Agarwal

ML @ CMU
Search Relevance @ Shopify
GPU Systems & MoE Research

Aarush Agarwal
Pittsburgh, PA
01Systems / Intelligence

Research

CMU Language Technologies Institute

Jan 2026 – Present

CMU Language Technologies Institute

Researching dynamic Mixture-of-Experts architectures under Chenyan Xiong, developing adaptive strategies that expand model capacity on out-of-distribution data while mitigating reasoning degradation in continual pretraining.

Designing autonomous research agents that propose, execute, and evaluate model-adaptation experiments across tasks and modalities including vision, clinical, and financial time-series data, iteratively committing variants that improve reasoning and task-benchmark performance over dense backbones.

CERN

Aug 2024 – Oct 2025

CMU Cosmology Laboratory & CERN

CUDA Researcher

First-authored the FastGraph paper and developed PCA-subspace pruning for exact GPU k-nearest-neighbor graph construction. The evaluated system achieves up to 41× speedups over FAISS-GPU on a 5-million-point, 8-dimensional HGCAL-derived workload with 40 neighbors on an A100 80 GB GPU.

Built on Jan Kieseler’s initial kernels: rewrote selected kernels for compile-time specialization, made targeted optimizations, and developed the PCA enhancement. Rewrote the manuscript from the ground up after an initial draft by Jan.

02Industry / Practice

Experience

Machine-learning engineering, open-source ML systems, and investment work across commerce, fraud, search, and early-stage technology.

SGLang

November 2026 – Present

SGLang & FlashInfer

Open-Source ML Systems Contributor

FP8 GEMM Autotuning: Corrected L2 cache contamination and latency measurement errors in SGLang's block-wise GEMM tuner, improving how it selects kernel configurations. The resulting configuration achieved 7.7% lower kernel latency on a controlled RTX 4090 workload.

GPU Memory Safety: Diagnosed FlashInfer's Gated DeltaNet prefill memory corruption under CUDA MPS and developed a fix that prevents out-of-bounds tensor-map writes under reduced GPU allocations. Built a cached device-level probe to size workspaces from hardware SM-ID bounds while preserving CUDA graph capture compatibility.

Persistent Kernel Caching: Implemented CuTe-DSL disk caching for FlashInfer's MXFP8 GEMM runners, allowing compiled kernels to survive process restarts. Reduced first-call time from 330 ms to 40 ms on a Blackwell warp split-K workload. Added architecture-specific cache keys, source-based invalidation, and automatic recompilation when cached kernels cannot be reused.

MoE Tuner Reliability: Fixed nullable quantization metadata parsing in SGLang's Triton MoE tuner, preventing configuration errors and correctly distinguishing channel, tensor, block, and grouped quantization recipes.

Shopify

Jun 2026 – Present

Shopify

Machine Learning Engineer Intern

Merchant-Aware Ranking: Designed and implemented a reusable auxiliary-loss framework for Shopify's search reranker. Used it to develop a merchant-aware training objective for merchant-intent queries, improving Shopify's Merchant Plus relevance metric by 11 percentage points. The framework enabled future work on incorporating signals such as user engagement and high-quality clicks.

Query Rewriting & Inference: Pruned and distilled Qwen3.5 for query rewriting, matching the search relevance of Shopify's production model and serving faster across nearly every tested load and latency percentile on production NVIDIA B200 MIG slices. Cut p50 latency by 23% against Shopify's tuned in-house inference stack by optimizing TurboMind with CUDA Graphs, programmatic dependent launch (PDL), custom INT8 decode kernels, and FP8 prefill.

Feedback-Driven Training: Built a pipeline that mines failed query rewrites from production and turns verified failures into preference pairs for targeted direct preference optimization (DPO).

Felicis

January 2026 – June 2026

Felicis

Venture Fellow

Co-organized Felicis × CMU VentureHacks, selecting 53 builders from 150+ applicants for an eight-hour build sprint that produced 20 project demos; awarded $10K+ in prizes and hosted an AMA with a Skild AI founding researcher.

AI Venture Research: Conducted startup diligence and market research on emerging AI companies and technical market shifts alongside the Felicis team.

Shopify

May 2025 – Aug 2025

Shopify

Machine Learning Engineer Intern

Fraud Detection: Improved buyer-fraud detection accuracy by 3% and reduced training iteration time by 70% through dimensionality reduction, importance-based feature pruning, and BigQuery/Dataflow + Vertex AI pipeline rebuilds.

AI Agent Network: Co-filed a patent for a distributed multi-agent system that decomposes tasks with a Neo4j dependency graph and executes subtasks across specialized agents in parallel.

Sequence Modeling: Designed transformer-based fraud models with embeddings and temporal attention over transaction sequences.

03Selected Work

Projects

A focused set of systems that connect models, hardware, and interfaces to real outcomes.

3rd Place · Hackberry Pi

Among 250 participants

AutoReflex

A sub-30ms hardware-in-the-loop aim assist that turns visual detections into physical corrections. A Jetson Orin Nano tracks targets at 100fps and streams coordinates over UDP to a Raspberry Pi, which drives a servo and optional TENS stimulation through a 1kHz PID controller.

AutoReflex — image 1 of 4
Image 01 of 04

< 30 ms

end-to-end latency

ProVision

AI Moneyball for Table Tennis. Transforms match film and tournament footage captured by 2D camera into actionable insights backed by tracked ball trajectories, 3D biomechanics, and Moneyball-grade analytics in minutes.

Image 01 of 03

2D → 3D

analysis pipeline

2nd Place, Healthcare · NexHacks

Among 1,500+ hackers

Healthier

An AI-powered care layer that makes senior care observable through multimodal sensing, real-time feedback, and clinician-facing risk insights.

Healthier — image 1 of 3
Image 01 of 03

Multimodal

home-to-clinic signals

1st Place, YC + Fetch.AI · Best Startup · CalHacks 12.0

Among 2,000+ participants

Haven

A multi-agent hospital intelligence platform that coordinates autonomous AI agents for patient monitoring and clinical decision support.

Haven — image 1 of 3
Image 01 of 03

Multi-agent

hospital intelligence