I train and evaluate language models under tight compute budgets and analyze GPU execution down to SASS. I am a graduate computer science student at Georgia Tech OMSCS, based in San Francisco.

• I submitted a 1.0805 BPB three-seed result to OpenAI's Parameter Golf challenge as pull request #2005, which is open and unmerged, and the full project write-up lives in DL-Team-Proposal.
• I keep a running study repository of AI and GPU papers with my own notes at James-neural-network-study, most recently adding interactive search and graph visualizations for CS 6601 Artificial Intelligence.
• I am building SF-city, a Godot and Python simulation of San Francisco where synthetic residents commute between imported real streets and buildings, with optional Rust helpers accelerating selected operations.

• DL-Team-Proposal is the CS 7643 team project that trained a language model to fit inside a 16 MB compressed artifact on 8xH100 GPUs with PyTorch, reaching 1.0805 BPB against an official naive baseline of 1.2244.
• AI-CDS-Disease-Diagnosis-Reproduction-Alpha reproduces a 2022 IEEE Access clinical decision support paper, fixing a patient-level data leakage bug and matching the 21 GB BioSentVec baseline on every retrieval metric with a 416 MB Bio_ClinicalBERT encoder.
• MongoDB-Agentic-context-window benchmarks OpenAI models on the BABILong needle-in-a-haystack task across context lengths, where gpt-4.1 held 91% accuracy at 64k context against 87.5% for gpt-4o-mini while costing about 40 times more.

• backprop-by-hand is the public notes and experiments companion to my private CS 7643 coursework, covering hand-derived backward passes from a two-layer MLP through Transformers and DDPMs, and reporting 97.8% MNIST test accuracy. No assignment solution code is included.
• gpu-showcase turns two CS 8803 GPU projects into public animations, including a SASS branch-divergence analyzer that flagged 136 divergent branches across 20 compiled CUDA kernels. The tensor-core simulator extension matched the course reference exactly but ran slower on every half-precision benchmark. Live: animations
• Robotic-AI-A-start-Search-and-Drone-SLAM is a write-up and demo-GIF showcase of three Artificial Intelligence for Robotics projects, covering particle-filter spaceship localization, online-graph SLAM for a drone, and A* pathfinding. The source stays private under the Georgia Tech honor code.

• WeddingAI is a solo Stanford x DeepMind hackathon submission that turns a phone video of an empty wedding venue into a Gemini-designed theme and builds toward a 3D Gaussian-splat reconstruction of the room. Running the reconstruction by hand on a rented RTX 5090 reached PSNR 25.29, and the deployed app serves that scene as a labelled sample. Live: app
• parameter-golf holds my entries to OpenAI's Parameter Golf challenge for the best language model that fits in 16 MB, where my headwise gated attention submission reports a self-measured three-seed mean of 1.0805 BPB. A later submission appeared to reach 1.0066 BPB until I found its byte accounting inflated and retracted it, and the corrected run scored 1.0972.
• watch-and-learn is a team entry to the MongoDB 2026 hackathon where a user watches and interacts with a Gemini agent driving a browser, and my VoyageAI embedding setup and multi-agent competitor-analysis feature were merged upstream.

• Rust-learning is an 80-lesson Rust curriculum I built for myself, where every lesson ships an example binary, a stubbed exercise and an integration test, and two Rust tools render a progress dashboard from the results.
• splat-service is a photos-in, 3D-scene-out service with a Rust and Axum backend that owns the job state machine, a Next.js viewer, and a RunPod GPU worker for COLMAP and Gaussian splat training. The GPU worker is still a Phase 1 stub, so the flow runs end to end today in a zero-credential mock mode.
• Ada is an open-source AI hardware engineering agent that turns a plain-language board description into a placed and routed KiCad layout, and I contributed eleven merged pull requests, including the Google ADK pipeline rewrite.
• jamesemerson112.github.io is my portfolio site, a Svelte application that scans my GitHub repositories with scc on a scheduled GitHub Actions run and renders searchable project cards and spider charts from the generated metrics. Live: site

I also write Rust for private client work, including a deployed Las Vegas nail salon scheduler and an orchard farm booking system still in progress, both built on Axum and SQLx against PostgreSQL.

Champion of Basketball 5v5 Intramural Greek Life at University of Nevada, Las Vegas

Experienced player in Squad (50 vs. 50 simulated battlefield). Squad lead at over 50+ matches. Commander at 10+ matches

Ex-President of Shine City volunteer. Vice-President of 25+ club members. Proud member of UNLV Buddies and ACM.

Prompt Engineer since ChatGPT 3.0

Email me at james.emerson.vo.2503@gmail.com if you want to talk about any of this.