Nathaniel Black

A Rust driver that turns a Hall-effect keyboard into an analog gamepad. It polls per-key travel depth off the board’s QMK raw-HID channel and maps W/A/S/D depth onto a virtual stick, so a light press walks and a hard press runs. I worked the board’s undocumented depth protocol out from live captures, then shipped two output backends behind a shared HID core, Linux uinput and Windows ViGEm, and confirmed on real hardware that the axis values the OS reports back match what the driver emits.

• Rust
• HID
• uinput
• ViGEm
• egui

I ported Quake III’s bg_pmove.c movement physics to Python from the C source, reproducing the projection-based acceleration rule that makes the exploit possible, and wrapped it as a Gymnasium environment. A PPO agent rewarded only on horizontal speed rediscovered bunnyhopping on its own, holding around 650 units per second and peaking near 970 against the engine’s 320 cap. To show the speed came from the ported physics rather than an artifact of the network, I reproduced the same exploit with hand-written inputs. 91 pytest cases cover the physics, the environment API, and the demo codec that replays trained runs inside the real game.

• Python
• Gymnasium
• Stable-Baselines3
• NumPy
• pytest

Multi-class classification on 145,394 NOAA Storm Events records from 2023 and 2024, predicting the ten event types that cover 81% of the database. Tuned XGBoost beat a swept TabNet at 0.900 against 0.851 macro-F1, and I confirmed the gap with a paired bootstrap and a McNemar test rather than leaving it at one number. The result I found most interesting was a calibration audit: TabNet’s lower expected calibration error turned out to come from underconfidence, not accuracy, and one-parameter temperature scaling cut XGBoost’s ECE from 0.025 to 0.007 without changing a single prediction.

• Python
• XGBoost
• TabNet
• pandas
• scikit-learn

A website planning tool built across bi-weekly sprints. We decomposed 25 user stories into eight epics on a GitHub Projects backlog and closed 11 of 11 committed stories in Sprint 1, running the full Scrum cycle around it. I owned the project management epic and designed the data-access layer behind it, routing all persistence through an async service API so components never touch storage directly and swapping in a real backend changes one module.

• TypeScript
• React
• Vite
• Jest
• GitHub Actions

A desktop configuration editor and game launcher for Quake III CPMA, with active users in the competitive community. It reads existing config files, converts console commands into typed in-memory settings, and regenerates valid Quake III output on save, which replaces editing those files by hand. First-run setup downloads and unpacks the game assets or links an existing install, and the whole thing ships as a standalone PyInstaller executable so it installs in one click on Windows without Python.

• Python
• Tkinter
• PyInstaller

Ported a two-hidden-layer MNIST network from serial C to CUDA, cutting a five-epoch training run from roughly 390 seconds on CPU to 2.2 seconds on an NVIDIA V100 while holding classification accuracy. I wrote the non-GEMM kernels by hand, including batched bias-add with fused ReLU and a per-row softmax with shared-memory reductions, and handed the three matmul-shaped operations to cuBLAS instead of hand-rolling them. Built and run on the university’s DARWIN cluster through Slurm.

• C
• CUDA
• cuBLAS
• Slurm

A 360-degree campus exploration game for incoming UD students. You get a panorama, then drop a pin on a Leaflet map where you think it was taken, with a Supabase-backed real-time leaderboard holding persistent scores and round history. I captured and processed the 50+ panoramic images myself with a 360-degree camera, owning that pipeline alongside the frontend and backend work.

• TypeScript
• React
• Three.js
• Node.js
• Supabase
• Leaflet

A CLI research system that replaces one-shot LLM summarization with explicit depth controls, pluggable search backends, and a fixed scoring rubric, so two runs are actually comparable. Candidate generation is separated from evaluation, and branches run in parallel across subagents that return structured mini-results for the orchestrator to merge, which buys breadth without collapsing onto the first plausible answer. A persistent evidence store keeps every claim’s source link, so a score can be re-checked later instead of regenerated from scratch.

• Python
• LLM agents
• pytest

An Apify actor that turns the Houzz professional directory into a clean lead dataset, using a two-tier crawl that reads structured JSON-LD for base records and optionally enriches each profile with ratings, reviews, and licensing. Deduplication keys on the stable Houzz profile id, dropping 546 overlapping and sponsored duplicates to reach 500 unique records with no collisions. Rather than claim coverage, I measured per-field fill rates over that real run, documented 98% phone capture, and kept nulls as signal.

• Python
• Apify
• BeautifulSoup
• pytest

Image segmentation by k-means, implemented three ways, from a naive loop up to a vectorized NumPy version benchmarked against scikit-learn. Running all three on the same images was the point: it puts a number on what unvectorized Python actually costs.

• Python
• NumPy
• scikit-learn