Hackathons

Built FRTC, an autonomous fraud-ring investigator for bank analysts. It analyzes 90 days of transactions, coordinates six specialist agents plus a Skeptic through a shared Cognee memory graph, and streams the investigation live to a graph UI.

Built AI Blox, a visual, drag-and-drop AI engineering tool inspired by Scratch. Snap together 500+ building blocks for LoRA fine-tuning, RAG pipelines, and multi-agent orchestration, then generate production-ready code instantly.

Placed 2nd in a live money-on-the-line prediction tournament, turning 50 into ~86 USDC (+72%) across 4 live rounds. Ranked ~90 anonymized futures by 20-day returns and ran the full pipeline: walk-forward tests, model training, staking scripts, and live execution. Gains came from targets, not algorithms: a 3-variant LightGBM blend with the benchmark, an Azurki auxiliary-target hedge, plus a distill and MLP for diversity. Cut live underperformers and moved stake onto the core models once they beat the benchmark.

Built AI Security Analyst, an evidence-first vendor-security questionnaire agent for the Regodit track. It parsed policies, SOC 2 reports, and access reviews into sourced claims, filled a 66-question buyer workbook without inventing answers, then applied the buyer's own risk tables so the vendor sees predicted rating, escalations, and a fix-first list before sending.

Built ClaimAid, an AI-powered wound-care billing triage tool for post-acute care facilities. Ingested 300 synthetic patients from a rate-limited PointClickCare API, extracted wound fields from SOAP, prose, multi-wound, and narrative notes, then routed each case through a three-gate triage engine into auto-accept, flag-for-review, or reject, cutting the biller queue by 66%.

Built Denial Rescue: RossMD, a seven-agent healthcare-regulatory legal team in software. Feed it a payer denial, a HIPAA breach, or an OIG subpoena and it issue-spots the exposure, researches it against a ~1,900-authority corpus of federal health law, war-games the opposing side, and drafts cited attorney work product — supervised and streamed live in ~90 seconds.

Built NYU Pal, an agentic AI platform unifying a dozen scattered NYU student tools around three primitives — Space, Items, and Knowledge. Ships eight features in one MVP: study-space finder, marketplace exchange, peer mentoring with resume matching, live printer tracker, activity-partner matcher, community notes, sublet listings, and an AI professor lookup — all routed through Claude Sonnet 4.6 with tool calling.