AWS AUTA Hackathon Winner · GSoC 2026 · Graduate Researcher, UMD

Shipping ML infrastructure from notebook to Redshift scale.

AI/ML Software Engineer II with 5+ years across healthcare, automotive and fintech. Recently: vector search prototypes on Redshift, LLM platforms on 2.3M+ patient profiles, and HPC I/O research at UMD with Alan Sussman.

View Projects Get in touch
5+
Years in AI/ML
120 PB
Data engineered
2.3M+
Patient profiles
Jay Guwalani - professional headshot
Google Summer of Code 2026 · New recognition

Selected as a GSoC 2026 contributor with The Linux Foundation.

Working on the Accord Project's APAP / MCP server - shared service layer refactor, four-tier testing and multi-client documentation. 175 hours over 12 weeks, mentored by Niall Roche and Dan Selman.

Idea #4 · APAP & MCP POC: 53 tests · 98.55% cov 5 prior PRs to Accord
Read the case study
GSoC 2026 - The Linux Foundation - Accord Project - APAP and MCP Server - Contributor Jay Guwalani
AWS · Software Engineer (ML) Intern · Summer 2026

Shipped ML infrastructure on AWS Redshift for AI and LLM-based use cases.

Completed internship with the Redshift ML team designing reliable, scalable systems for production AI. Won the Amazon AUTA AI Hackathon, org-wide top winner. Key learning: at scale, the hard part isn't the model - it's everything around it.

Recent highlights

Selected case studies

Production systems built end-to-end, from data pipelines to deployed APIs and dashboards.

All projects
Enterprise LLM evals schematic
Oct 2026 · LLM Evaluation

Three Evaluations Missing From The Public Corpus

Twenty enterprise scenarios in openai/evals format plus three novel evals the public corpus does not cover: tool-use trajectory, streaming-cancel JSON validity, and PII leakage through tool arguments. Cross-model: Sonnet 4.6, Opus 4.7, GPT-4.X.

20 scenarios 3 novel evals Cross-model
Python Anthropic SDK openai/evals Prompt caching
Read case study
Agent benchmark schematic
Sep 2026 · Benchmarks

Replit Agent Bench: Head-to-Head on 7 Realistic Dev Tasks

Reproducible benchmark comparing Replit Agent, Cursor paired with Claude Sonnet, and the raw Claude API across seven real-world development tasks. One command to run, pinned model versions, open scoring rubric.

3 agents · 7 tasks make bench Quarterly reruns
Python pytest Anthropic SDK Docker
Read case study
Comet-shaped browser agent schematic
Sep 2026 · Browser Agents

Comet-Shaped Web Agent: Payload-Classified Permissions

Browser agent on Playwright and the Chrome DevTools Protocol, driven by Claude Sonnet 4.6 with tool use. Each proposed action classified by payload content before dispatch, so permission scope depends on what the agent is about to do, not just where.

Sonnet 4.6 Playwright + CDP Payload tiers
TypeScript Claude Sonnet 4.6 Playwright CDP
Read case study