Skip to content
View Arunjay4213's full-sized avatar
3 followers · 0 following

Achievements

Achievement: StarstruckAchievement: YOLOAchievement: Pull Shark

Achievements

Achievement: StarstruckAchievement: YOLOAchievement: Pull Shark

Highlights

  • Pro

Block or report Arunjay4213

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
Arunjay4213/README.md

Arunjay Agrawal

CS, Mathematics, and Data Science triple major at UW-Madison, graduating May 2028. I like problems that need both the systems layer and the ML layer: backends, data pipelines, GPU kernels, and the RAG, agents, and LLM tooling on top. I'm looking for a Summer 2027 SWE or AI engineering internship.

Portfolio: arunjay.dev

Work in private repositories

Most of my job and lab work is in employer or university repositories, so I describe it here instead of linking it.

  • Wisconsin Surgical Outcomes Research Program (WiSOR), Software Engineer, 2026 - present. A RAG system for thyroid cancer patients that checks each claim against curated clinical sources. It is evaluated against frontier models on a clinician-rated benchmark.
  • UW Tech Exploration Lab, Software Engineer Intern, 2026 - present. A bilingual English/Spanish veterinary voice agent in Go for a 10-veterinarian dairy farm pilot, and the lab's FastAPI and Next.js management dashboard.
  • UW-Madison Data Science Institute, Undergraduate Research Fellow, 2026 - present. Built the multi-speaker audio transcription tool for a qualitative research platform, and gave its research assistant sandboxed SQL access to an ontology knowledge graph.
  • RIA Advisory, AI Engineer Intern, summer 2026. An agentic system that turns unstructured client input into cited business requirements documents. It uses GraphRAG over Qdrant and Neo4j and a LangGraph pipeline with a critic loop.
  • LogicPatterns, Software Engineer Intern, summer 2025. An AI agent that turns natural-language prompts into live software architecture diagrams.

Public projects

  • race-attention-kernel - fused CUDA kernels for RACE Attention (ICLR 2026), built at the CDIS & OpenAI Summer AI Lab. The reference causal path runs a 40 GB A100 out of memory at 131K tokens. The fused kernel runs a 2M-token causal forward in 14.7 ms on one H100, 165x faster than the chunked PyTorch baseline.
  • jobcrawl - watches company career boards at the source and emails a scored digest of new internship postings. It has run unattended since July 2026: 7,358 runs, 5,894 career boards on 14 applicant tracking systems plus company APIs, and about 872 million postings fetched. ETag conditional requests and per-host throttling keep it polite and cheap, dedupe catches reposts, and it notices when a posting closes. An LLM scores whatever gets past a regex prefilter.
  • mosh-rs - a Rust port of the mobile shell (mosh). It is wire-compatible with mosh 1.4 servers, adds a native Windows client and real scrollback, and is published on crates.io (cargo install mosh). It was built with AI coding agents under a written plan, and the plan, status log, and reviews are all in the repo.
  • Tripwire - tests whether a tool-using AI agent can be hijacked by prompt injection before it ships. It runs 16 attacks, and plain Python (no LLM judge) checks whether the agent leaked a planted secret. Published on PyPI as tripwire-eval.
  • Cortexa - observability for AI agent memory. It records which memories each response used, scores their influence, and traces every memory back to the interaction that created it.
  • Wattbot - hybrid BM25 and dense retrieval for a UW-Madison Kaggle competition, 87% accuracy on 282 test questions.
  • CustomShell - a POSIX shell in C11 with pipelines, redirection, and job control.
  • Monday - vision-guided assisted grasping for people with upper motor neuron injury, using voice, a camera, a Claude grip planner, and electrical muscle stimulation. Built at StarkHacks 2026.
  • AI_Interviewer - an AI mock interviewer that also reads your body language from the webcam. 3rd place at MadData 2025.
  • claude-mods - four Claude Code mods written in TypeScript. They show context window use, plan quota burn rate, and per-turn cost on screen, and one stops the session past a budget you set.
  • shipwright - an agent that builds a web app from one line of English, deploys it, tests it in a real browser, and fixes it until the tests pass.

Contact

Pinned Loading

  1. race-attention-kernel race-attention-kernel Public

    Fused CUDA kernels for RACE attention (ICLR 2026). A 2M-token causal forward takes 14.7 ms on one H100, 165x faster than the PyTorch baseline.

    Python

  2. Cortexa Cortexa Public

    Observability for AI agent memory. Records which stored memories each response used and traces every memory back to where it came from.

    Python

  3. Tripwire Tripwire Public

    Checks whether a tool-using AI agent can be pushed by prompt injection into leaking a secret, before you ship it. On PyPI as tripwire-eval.

    Python 3

  4. mosh-rs mosh-rs Public

    mosh, the mobile shell, rewritten in Rust. Works with existing mosh-server 1.4 and adds a native Windows client and scrollback.

    Rust 27 2

  5. Wattbot Wattbot Public

    Hybrid RAG for the UW-Madison Environmental AI Kaggle competition. 87% accuracy on 282 test questions with only 40 labelled examples.

    Python

  6. CustomShell CustomShell Public

    A POSIX shell in C with pipes, redirection, job control and builtins.

    C 1