Aarush
Agarwal

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.

Merchant-Aware Ranking: Designed and implemented an auxiliary merchant-aware training objective for merchant-intent queries. Improved Shopify's Merchant Plus relevance metric by 10% by helping the ranker surface a merchant's first-party catalog ahead of third-party resellers.

Feedback-Driven Training: Built a feedback system that samples production search queries, identifies losing rewrites against reference judgments, and uses those failure cases for targeted direct preference optimization (DPO) of the query-rewriting model.

Training & Compression: Fine-tuned Qwen3.5-0.8B for query rewriting, then compressed it through structured layer pruning (24 → 16 layers) and reverse-BPE vocabulary reduction (248K → 65K tokens). Matched production term F1 through knowledge distillation.

Serving Optimization: Optimized TurboMind serving for the compressed Qwen3.5 model with CUDA graphs, custom INT8 decode kernels, and FP8 prefill. Cut benchmark latency by approximately 20% versus Shopify's existing internal query-optimization engines on B200 hardware, matching or beating production LFM2 latency across percentiles and evaluated query categories.

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.