Applied Machine Learning Engineer#
Hi, I’m Arjun, an Applied Machine Learning Engineer with 14+ years of building and shipping ML systems. I work across the whole stack: data and sampling, modelling, distributed training, and low-latency C++ / Python inference at scale.
I’m currently a Senior Applied ML Engineer at Shopify, working on merchant foundation models.
Machine Learning Proficiency#
Areas I’ve worked in across the full project lifecycle:
- Data & sampling: unbiased datasets via feature-based sampling, hard-negative mining, and synthetic data that matches the real distribution.
- Modelling:
Convolutional,Recurrent, andTransformerarchitectures; contrastive learning (SimCLR,BYOL,SimSiam); and sequential models. - Distributed training: GPU clusters over
OpenMPI + RoCE,Torch RPC, andPyTorch Lightning. - Calibration:
Focal Loss, label smoothing, isotonic regression, and Platt scaling. - Optimization: quantization, pruning, distillation, and graph fusing with
ONNX,TorchDynamo, andTVM. - Inference: scalable, low-latency
C++serving (ONNX,Drogon,Triton) onk8s.
Recent Explorations#
Some things I’ve built recently, and what each one looks at:
- imba-chess: a high-Elo chess bot that ports
HSTUsequential transformers from recommender systems to chess, with policy imitation from Lichess games, aWDLvalue head, and depth-2 policy-pruned minimax search. It looks at how far sequence models transfer from recsys to strategic games. Build log: Building a Chess Bot with HSTU. - muvfde: fixed-dimensional embeddings for multi-vector representations, based on Google Research’s
MUVERA. TheC++core is exposed to Python viananobindand published on PyPI. It trades model size for retrieval latency using a fixed-dimensional encoding. - Comparing Online Hyperopts: a benchmark of online hyper-parameter optimization methods on sampling efficiency, latency, and ease of implementation.
Domain Expertise#
Problems I’ve shipped machine learning for:
- Search & Ranking, Cold-start Recommendations
- Constraint-based Optimization
- Speech Processing (Speech-to-Text, Text-to-Speech)
- Computer Vision (Segmentation, Classification, OCR)
- Natural Language Processing (Document QA, Classification, Entity Recognition)
- Contrastive Learning methods