Work

Clinical data collection for ABA therapists — the people who teach children with autism through structured repetition. Therapists record trials on tablets in noisy classrooms, so every touch target is 140px wide and every interaction forgives imprecision. The mastery engine watches the data and tells the therapist when a child is ready to move forward.

Practice management for four home health agencies in New York. Therapists document visits on their phones. Supervisors cosign remotely. The system tracks sixty credential types and locks anyone out the instant a critical one expires — before an auditor ever has to ask.

Time tracking built in two days because the agency needed it yesterday. Employees clock in on a wall-mounted iPad. Managers see who’s working in real time. Timesheets generate themselves. Usage grew 2.4x in fourteen weeks because it just works.

Case management for Early Intervention — the pipeline from referral to active caseload. Kanban intake, child records with ICD-10 codes, sibling linking, referral staging, and consent document storage. Auto-saves on blur so nothing is lost mid-workflow.

Every five minutes, this system reads the agency’s inbox, understands the referral buried inside each email, and writes structured data to a spreadsheet. What used to take someone four hours a day now takes fifteen minutes of review.

A clinical document validator that runs a 3.8-billion-parameter AI model entirely in the browser. No data leaves the device — ever. Therapists upload PDFs and the system checks dates, terminology, contacts, and template compliance without a single byte touching the cloud.

My personal AI infrastructure — thirty-plus custom skills for research, code review, self-optimization, and investigative journalism. A system with memory, with specialized capabilities, and with a feedback loop that genuinely improves the more I use it.

A workout tracker you use with your voice. Tap the mic, say what you did, see it logged in under five seconds. All transcription happens on-device — zero cloud cost, zero data exposure. Detects personal records automatically and generates shareable cards.

Type your address, see every official who represents you — from city council to Congress. Track bills across fifty states. Follow the money through campaign finance, PAC contributions, and lobbying disclosures.

A ranking system for local BJJ academies, because knowing who’s actually the best on the mat shouldn’t require a tournament. Glicko-2 ratings update after every match. Every academy keeps its own leaderboard.

Interactive visualization of James Webb Space Telescope data — the light that passes through exoplanet atmospheres and tells us what they’re made of. Transmission spectra, time series heatmaps, molecular overlays for water, carbon dioxide, methane, sulfur dioxide.

A multi-threaded game server in C where players hunt each other through a maze with lasers. Thread-safe, memory-leak-free, built from raw TCP sockets and POSIX primitives. The kind of project that teaches you what the machine is actually doing.

A custom malloc implementation — the thing every C programmer should build once to understand what happens when you ask for memory. Segregated free lists, coalescing, alignment, and corruption detection through header obfuscation.

A job scheduler built entirely from POSIX system calls — fork, exec, pipe, dup2, and signal handlers. Process creation, suspension, resumption, and cleanup, all managed by hand.

A Protocol Buffers parser built without any protobuf library. Varint decoding, zig-zag encoding, delta compression, zlib decompression — all from scratch. The best way to understand a format is to read it byte by byte.

A passive network sniffer for my offensive security course — captures HTTP, TLS, and DNS traffic, identifies automation tools, and flags internal infrastructure hostnames. The kind of tool that teaches you how much a network reveals about itself.

An HTTP server and caching proxy built from raw sockets — no frameworks, no libraries, just the RFC and a socket. Handles concurrent requests and caches responses.

A batch audio transcriber running Whisper models natively on Apple Silicon through the MLX framework. Zero cloud dependency — everything stays on the machine. I use it daily for meetings, podcasts, and voice notes.

Can you predict how much a movie will make from its metadata alone? With a hundred engineered features and the right model, you get surprisingly close. Random Forest with five-fold cross-validation, R² = 0.77.

A year of five-minute weather snapshots mapped against residential energy consumption. The question: can weather predict how much electricity a home will use? Linear regression for daily demand, logistic regression for peak-load classification, appliance-level profiling.