Projects
AlignEd
Two-sided AI education platform where teachers manage classrooms and assignments, and students turn their course materials into shareable study artifacts (flashcards, quizzes, LaTeX problem sets, podcasts). Sole engineer, built as Research Assistant at Columbia.
Debt Vulture
Bankruptcy intelligence platform for distressed debt investors. Monitors PACER filings in real time, extracts structured data from legal documents, and maps affected public companies. Includes a multi-agent simulation engine where AI agents carry persistent memory, form insights through reflection, and negotiate under enforced information asymmetry. Domain-agnostic: same architecture runs bankruptcy, geopolitical, and M&A scenarios. First paying customer is a hedge fund.
River
Physics-based spatial task scheduler where tasks are organic blobs drifting in a current. Claude Code controllable via MCP.
NLP Hallucination Mitigation
Three-stage pipeline (CoT distillation, abstention fine-tuning, RLVF) reducing LLM hallucinations in closed-book QA. 97.5% precision in hallucination detection; Selective F1 improved from 0.50 to 0.70.
NASA SUITS
Scaled team to 20 engineers; built AR astronaut guidance system tested at Johnson Space Center with NASA engineers.
More Projects
Josh-OS
Click to view
Puriphico
Click to view
ArcaTracker
Click to view
Don't Fret!
Click to view
Claude Code Setup
I tutor non-technical professionals on AI. The thing I keep learning: Claude Code ships like a MacBook Pro with no software installed. Writing clear instructions, documenting failure modes, and building repeatable workflows-- that's the software. Here's mine!
The orchestration skills get the most use. I usually have 5+ Claude Code sessions running at once, and Jarvis, Manager, and Consult are how I keep track of what they're all doing. Jarvis reads my recent sessions, git commits, and vault changes so the current session knows what I've been up to without me explaining it. Manager lets any session monitor, message, or kill the others. Consult lets one session ask another session a question without me switching windows. For overnight builds I hand a spec to the autonomous-vm skill and check in the morning.
The corrections skills exist because Claude Code has habits that drive me crazy. It over-engineers, hedges when I want a direct answer, and writes prose that sounds like AI. So I built /vent to log the frustration while it's fresh, /therapy to cluster the gripes and design actual fixes (CLAUDE.md rules, hook scripts), /human-writing to kill AI patterns in anything it writes for me, and /second-opinion to break tunnel vision by querying GPT or Gemini for an independent take.
Easiest way to install: Download All, extract the zip, and point Claude Code at it. The zip includes a README with setup instructions that Claude Code will read and follow — it'll ask you where to put things and handle the rest.
Resume
Loading resume...