Smit Kiri

DART Jobs had been running successfully for two years, but its underlying architecture had limits that constrained how ML teams could use it. I led the ground-up redesign: defining the system architecture, evaluating tradeoffs, and serving as technical lead through execution. DART Jobs V2 drove a 7x increase in modeling runs and doubled the number of new users on the platform, unlocking a new generation of ML-powered products that weren't feasible before.

By 2024, DART had grown from infrastructure into institution — 40 systems across 6 teams running on the platform. Teams that had spent months building custom deployment pipelines were instead shipping models in days. One team achieved a 10x speed-up in model training, another migrated ~10 models and shut down their legacy deployment system entirely, and a third deployed a suite of new compliance models that had no path to production before DART existed.

Distributed training and large-scale batch inference across hundreds of nodes wasn't possible with Klaviyo's existing infrastructure. I built DART Jobs, the platform that changed that — covering cluster provisioning, job scheduling, fault tolerance, and observability, all running on Ray and Kubernetes. DART Jobs now handles 350–450 ML jobs per day and has become the foundation for Klaviyo's most compute-intensive research and production workloads.

At Klaviyo, each ML team was building its own serving infrastructure from scratch. The process was slow, duplicated across teams, and pulled engineers away from model development. I designed and built DART Online, a shared inference platform on Ray Serve that standardized the path from trained model to production. Standing up a new serving endpoint went from a months-long engineering effort to a matter of days.

Named Entity Recognition and Relation Extraction from Electronic Health Records. Uses BioBERT for entity recognition and a BiLSTM-CRF architecture to extract relationships between medical entities like drugs, adverse events, and dosages from clinical text.

A reading comprehension based question-answering model for news articles. Built with BERT and DistilBERT transformer models in PyTorch, trained on the Microsoft NewsQA dataset.