Parth Parmar

Bachelor of Applied Science in Engineering

Minor: Computer Science

Amazon - New York, NY

• Played a key role in the beta launch of an advertiser-facing AI agent that analyzes campaign performance data and generates actionable insights, improving advertisers' ability to plan and optimize campaigns.
• Led development of an Ads AI agent framework on AWS Bedrock AgentCore and Strands SDK enabling multi-turn conversational agents with streaming responses, tool orchestration, MCP integrations, and memory persistence supporting short-term and long-term context with per-session and per-advertiser isolation.
• Architected and built a distributed synthetic data generation platform that produced millions of high-quality LLM training samples across thousands of advertisers, enabling faster post-training workflows and overcoming cold-start constraints with multi-step ReAct reasoning traces and tool use via teacher models.
• Developed a scalable LLM-as-Judge evaluation platform generating real-time reward signals for RL training and model evaluation pipelines, supporting high-throughput synchronous and asynchronous batch processing across thousands of evaluations.

Aryn.ai - Mountain View, CA

• Collaboratively developed an AI-powered open source document processing framework, Sycamore, for RAG and unstructured analytics using Ray.
• Led development of an MVP inference service using an in-house DETR model, enabling users to segment and partition PDFs efficiently.
• Spearheaded the design and development of the control plane for the managed offering of the Sycamore platform, enabling the company to attract early adopters.
• Regularly reviewed code and mentored junior developers, reinforcing engineering quality and continuous improvement.

Amazon Web Services - Palo Alto, CA

• Successfully launched a beta version of a managed service providing data storage optimization for consumer data lakes, improving efficiency and reducing cost.
• Worked in a cross-functional environment to launch ACID transaction support in S3 data lakes.
• Was an early team member of AWS Glue Elastic Views, a service that lets customers create materialized views across multiple data sources using familiar SQL without managing infrastructure.
• Led the design and development of on-host TLS termination for two major services to achieve encryption in transit.
• Collaborated with senior software engineers to design and develop a multi-tenant incremental view maintenance service using AWS technologies.

Flipp - Toronto, ON

• Assisted team leads and senior developers in creating the technical specs and architecture of a new digital publishing system.
• Built and tested several microservice features with TypeScript, Node, and Kafka, reducing flyer generation time from 12 hours to about 1 hour.

University of Toronto - Toronto, ON

• Used NLP to determine cause of death from written descriptions of a person's symptoms leading up to death.
• Created a machine learning pipeline in Python to investigate multiple algorithms and optimize classification accuracy.

• Co-authored "The Design of an LLM-powered Unstructured Analytics System" (arXiv:2409.00847) during my time at Aryn.ai for the Conference on Innovative Data Systems Research (CIDR).