iammusa.vercel.app · LinkedIn · mmusa2@wisc.edu · ORCID · Setup notes
I'm Musa, and I'm finishing my M.S. in Computer Sciences at UW-Madison this December. Most of my research has been on LLM security, prompt injection in particular, and since last fall I've been working with Prof. Somesh Jha on agentic knowledge-graph extraction from energy-regulatory rate cases.
Outside research I lead engineering at WUMI Health. I also TA the graduate Big Data Systems course, where I look after the programming projects and their autograders (before this semester I was head TA for Data Science Programming II).
| Project | What it is |
|---|---|
| distributed-lender-service | A C++20 gRPC service over MySQL, Parquet and HDFS, about 440k mortgage records. Its cache sits at replication 1 on purpose, so when a DataNode dies the service has to notice and rebuild it. There's a lock-free ring buffer and an epoll server in there too. |
| agent-distillation-lab | I distilled a 4B model's tool-use policy into a 0.6B one with LoRA. All of it ran on my laptop's 6 GB GPU. |
| triage-guard-agent | A LangGraph support agent. Refunds over $50 wait for a human. |
| artifact-conformance-toolkit | Checks that read a benchmark repo's docs, raw results and git history, and fail it when a published number can't be traced back to the records behind it. |
| kiln | I save a lot of reels and never open them again. Kiln reads them for me and turns them into notes I can search. A copy of my library is up at kiln-by-m.vercel.app. |
| internships | Work from my internships at DeepLearnHQ (2025) and SAUDCONSULT (2024), rebuilt on my own machine. |
| musa-portfolio | My personal site: Three.js, plus an AI twin you can ask about my work. |
Older: AugData (a benchmark dataset for multimodal models), customized_transport_layer (reliable transport over UDP, in Python) and reprod, the code behind the ransomware paper below. WUMI Health, resume-gauntlet and the knowledge-graph project are private.
Last spring a partner and I did a course project on indirect prompt injection in RAG, and I built most of the attacks and defenses. The system-prompt defense OWASP recommends backfired on us. In a paired test on llama3.2:3b it took one attack's success rate from 12% to 38%.
I've also compared protein language model embeddings as GraphSAGE features for protein-interaction link prediction, where the structure-aware SaProt model did best (AUPRC 0.838). And with a collaborator I built the first labeled topic-classification dataset for Punjabi written in Shahmukhi script.
Papers: Towards Reproducible Ransomware Analysis, ACM CSET 2023, co-author with SRI International. Extremism on Social Media: Lynching of Priyantha Kumara Diyawadana, IEEE/ACM ASONAM 2022, first author.
WUMI Health is a patient-owned health record for Pakistan, built in Flutter on Supabase with consent enforced by Postgres row-level security. We're seven phases into a thirteen-phase build, and seven clinics, two hospitals and three labs are on board. I work on it with three interns.
I build with Claude Code and have spent a lot of time on how it's set up. There's a write-up in the setup notes.
Undergrad was computer science at LUMS in Lahore. I was on the Dean's Honor List, ran the IEEE student chapter, and was project head for a UAV team that made the top 20 at TEKNOFEST. Fun fact: I love sports, travelling and CS, three things that are impossible to manage together, so you know I have good time management skills.
LANGUAGES
MACHINE LEARNING & DATA SCIENCE
LANGUAGE MODELS & AGENTS
DATA ENGINEERING
DATABASES & SEARCH
BACKEND & APIS
WEB, MOBILE & 3D
CLOUD & DEVOPS
AUTOMATION & SCRAPING
DESIGN & DOCS
Also used, without a logo above: OpenAI API · LlamaIndex · gRPC · Protocol Buffers · XGBoost · LightGBM · PEFT/LoRA · TRL · FAISS · ChromaDB · pgvector · DeepEval · Playwright · dbt · Great Expectations · Power BI · Matplotlib · seaborn · Altair · statsmodels · Groq API · Hugging Face Accelerate · React Three Fiber · Framer Motion · Detectron2 · Riverpod · Canva
The snake gets redrawn every night from my contribution graph.

