AA

Hi, I'm Arunjay

I study CS, Math, and Data Science at UW-Madison and build AI systems and the infrastructure underneath them: agent security testing, CUDA kernels, and retrieval pipelines.

Download resume View my work

arunjay101@gmail.com / GitHub / LinkedIn

About

Get to know me

I'm an undergrad at the University of Wisconsin-Madison studying Computer Science, Mathematics, and Data Science, graduating in May 2028.

Most of my work sits where AI meets systems engineering. I've built agentic pipelines that draft cited business documents for an enterprise healthcare-payer platform, a security harness that measures whether tool-using agents can be hijacked by prompt injection, and I'm currently writing a fused CUDA kernel for a linear attention mechanism. I care about the unglamorous parts that make these systems trustworthy: deterministic evaluation, throttling and retry logic, provenance, and tests that actually catch regressions.

Outside of internships and research I'm always building, usually with several projects in flight at once: open-source software, personal tools that automate bits of my day-to-day life, and hackathon builds. I compete in hackathons regularly, taking 3rd place at MadData and prototyping assistive hardware at StarkHacks.

PythonTypeScriptGoC CUDAPyTorchLangGraph FastAPIPostgreSQLDocker GCPAWS

Experience

Where I've worked

Jun 2026 - Aug 2026

AI Engineer Intern

RIA Advisory, Miami, FL

  • Built an agentic AI system that converts unstructured client inputs into cited business requirements documents for an enterprise healthcare-payer Oracle ORMB implementation, cutting consultant time-to-first-draft by 60%.
  • Architected a scoped GraphRAG retrieval layer over Qdrant and Neo4j indexing 4,000+ pages across 7 knowledge bases, and a multi-agent LangGraph pipeline with an LLM-critic loop that revises drafts until they pass faithfulness checks against cited sources.
Feb 2026 - Present

Software Engineer Intern

UW-Madison Tech Exploration Lab, Madison, WI

  • Built a bilingual Spanish/English veterinary AI voice agent in Go serving 150+ daily queries for a 10-veterinarian pilot across Wisconsin dairy farms, with strict non-diagnostic guardrails and a 50-case safety test suite.
  • Created a full-stack lab management dashboard (Python, JavaScript, PostgreSQL) on GCP, automating mentor matching and resource allocation across 30 active ventures.
Feb 2026 - Present

Undergraduate Research Fellow, CALL-ECL Project

UW-Madison Data Science Institute, Wallace Foundation $102M Initiative

  • Mapped and documented the architecture of a 120,000+ line production monorepo (React/TypeScript, Python/Flask), delivering formalized system documentation and refactoring legacy components to reduce technical debt.
  • Modernized the codebase powering an equity-centered leadership platform serving 8 urban school districts and about 500,000 students.
  • Designed pipelines that transform coded qualitative transcripts into social network analysis adjacency matrices, plus an interactive graph explorer (Cytoscape.js, D3.js, NetworkX) supporting up to 1,300 nodes.
May 2025 - Sep 2025

Software Engineer Intern

LogicPatterns, San Francisco, CA

  • Built a full-stack AI agent that converts natural language prompts into live software architecture diagrams, cutting diagram creation time from 30+ minutes to under 2 minutes.
  • Containerized the backend as FastAPI microservices with Docker and wrote Pytest integration tests covering inference, diagram synthesis, and frontend rendering.
  • Optimized client-side SVG rendering and built a JSON-to-React bridge that brought diagram render latency down to 1.3 seconds.
  • Wrote 3 web crawlers that indexed 1,000+ cloud-service icons across AWS, GCP, and Azure documentation, storing metadata in PostgreSQL (JSONB) for accurate diagram generation.
Expected May 2028

University of Wisconsin-Madison

B.S. Computer Science, B.S. Mathematics, B.S. Data Science. GPA 3.82/4.0.

Projects

Highlighted work

Tripwire

2026 Live on PyPI

Prompt-injection evaluation harness

Python, GitHub Actions CI, LangGraph/CrewAI adapters

Measures whether a tool-using AI agent can be hijacked by prompt injection before it ships: 13 attacks, deterministic pass/fail with no LLM judge, attack-success rates with confidence intervals, and plain-English hardening advice. Published as tripwire-eval with a 351-test suite and a CI release gate on attack success rate.

Terminal output of a tripwire run: an attack-success-rate table for three attacks, followed by security feedback explaining which attacks leaked and how to harden the agent

PyPI GitHub

RACE Attention Kernel

2026 In progress

GPU systems work, CDIS & OpenAI Summer AI Lab

CUDA, PyTorch, bf16

Fusing the causal forward pass of RACE Attention (ICLR 2026) into a single CUDA kernel. Traced the reference implementation and found its shipped CUDA kernels never run; the real bottleneck materializes a five-dimensional tensor that runs a 40 GB A100 out of memory at 131K tokens. The fused kernel removes it entirely, targeting 2M-token forwards on a single H100.

Log-log chart of peak GPU memory versus context length: the reference cumsum path at 284 KB per token hits out-of-memory at 131K tokens on a 40 GB A100, while the fused kernel at 4 KB per token reaches 2M tokens using 8.4 GB

GitHub

jobcrawl

2026 Running 24/7

Autonomous job-search pipeline

Python, SQLite, Claude API, systemd

Polls about 540 company career boards at the ATS source every 10 minutes, scores new postings against my profile with Claude, and publishes email digests plus an auto-generated tracker. Conditional fetches with ETags, per-host throttling, content-hash dedupe with repost detection, and absence-based closure detection. Runs unattended on an EC2 instance under a systemd timer. Source is private.

Architecture breakdown

See all projects

Contact

Get in touch

The fastest way to reach me is email.