← back to graph

Alex Guha

alexandersguha@gmail.comgithub.com/Alex-Guhaalexguha.comlinkedin.com/in/alexguha/

Education

Work Experience

Tempe, Arizona Aug 2024 – May 2026
Graduate Researcher · Arizona State University
  • Developed a framework that puts a measurable number on how many features a transformer's latent space can represent, extending the superposition hypothesis with a near-orthogonality tolerance estimate read from embedding geometry.
  • Built a model analysis pipeline featuring VRAM-aware slicing that tiles billion-element matrices to available memory and cached intermediate tensors for cheap re-runs across dozens of HuggingFace checkpoints.
Tempe, Arizona Aug 2023 – Aug 2024
Team Lead · RAG-based Q&A System for CS Courses (ASU Bachelor's Capstone)
  • Worked with a professor to design and implement a machine learning pipeline utilizing retrieval augmented generation and OpenAI's ChatGPT to answer students' questions.
  • Synthesized a complex text dataset from the various resources in multiple computer science classes at ASU.
  • Presented the project at an international Intel competition hosted in Shanghai and won a second place award.
  • Won an additional cash prize at a separate international competition.

Research

2026
Representational Capacity: Geometric Limits on Feature Representation in Transformer Language Models
  • A measurable proxy for near-orthogonality tolerance plus an adjusted JL bound gives a quantitative cap on features a transformer can represent.
  • Preprint: arXiv:2606.02765 · Code & paper source.

Projects

2026
WynnSolver · Build optimizer with closed-loop combo simulator on a ~10²⁰ search space
  • Designed a multithreaded heuristic search returning ranked builds in seconds across an otherwise-intractable space.
  • Pure JS, zero-backend; incremental updates and pre-allocated scratch maps keep memory usage low.
2026
Personal Website · Interactive 3D graph homepage - Rust → WebAssembly via wgpu
  • Built a 3D graph homepage where each node is a navigable sub-page, rendered against a real starmap; reusable wgpu crate handles rendering with no DOM/input coupling, and the wasm graph crate handles layout, input, and navigation.
  • Authored an autonomous boids flock that drifts around the graph - vectorized neighbor queries, separation/alignment/cohesion tuned for the 3D camera, and idle-time animation that doesn't fight user interaction.
2022 - Ongoing
Home Lab
  • OPNsense router and TrueNAS server hosting personal media, photos, a password manager, and custom Docker apps.
  • Spare desktop compute hosting game servers (mostly Minecraft) over WSL with wake-on-LAN.

Skills

Languages : Python, JavaScript, C++, HTML, CSS
ML / AI : PyTorch, HuggingFace, transformers, LLM Architecture, RAG
Tools & Infra : Git, Docker, Linux, GitHub Pages, TrueNAS, OPNsense
Interests : Rock climbing (collegiate nationals - men's intermediate), drone photography, cooking