M.S. Data Science (AI concentration), Fordham University. B.A. Economics, University of Virginia. Founding engineer shipping AI products end to end, applied AI across fintech and healthcare, from prototype to production for real users.
SPY/sector-hedged mean reversion + momentum strategy. Achieved 0.91 Information Ratio, 52% win rate, and 2.24 bips net PnL. Fully backtested pipeline with live signal generation.
37-factor risk model with Ledoit-Wolf shrinkage and CVXPY optimization. Achieves 27-29% vol reduction without capital gains realization. Semantic nearest-neighbor hedges via 768-1024d embeddings.
Flask + React + Flux AI + N8N + MediaPipe fallback pipeline. Demographics-aware prompt engineering, 99.2% uptime. Presented at SIT AI Appreciation Day to AWS and Big 4 reps.
Three-phase agentic workflow (Monitor → Analyze → Decide) for autonomous portfolio rebalancing. Evaluates 4 rebalancing scenarios with VaR, Sharpe, and beta scoring. FinBERT sentiment integration for regime-aware decisions.
End-to-end ML pipeline for factor-based equity return prediction using 1999–2025 Bloomberg factor data. Model bake-off across XGBoost, BayesianRidge, and adaptive blending with walk-forward backtesting and SHAP explainability.
Real-time risk analytics system across 3 portfolios and 20 tickers. MongoDB Atlas for historical VaR, Sharpe, and beta storage; Redis Cloud for sub-10ms live metric reads. Streamlit dashboard with configurable alert thresholds.
MCP server combining real-time sentiment analysis, live stock data, and paper trading execution. Gemini AI sentiment aggregated from news and Reddit. Intelligent caching reduces API calls by 60%+. Batch processes 5 tickers in under 5 seconds.
AI-powered academic advising system with graduation risk assessment, career-based course recommendations, prerequisite visualization, and post-graduation pathway planning. Early warning system flags at-risk students before late graduations occur.
Find paid clinical trials and get screened in one phone call. Candidates browse real recruiting studies compiled from ClinicalTrials.gov, call a live Twilio line, and a voice agent runs the study’s actual published eligibility criteria as a conversation — rejecting, cross-offering, or qualifying and booking a real coordinator callback inside the call. Dropped calls are called back and resume at the exact pending question. The LLM never decides eligibility: it only extracts structured answers, and deterministic code judges them against registry-verified criteria. Every spoken sentence is a code-owned template, the same verbatim-script property IRB-approved pre-screens run under.
Verified health goals with instant USDC rewards. Sponsor-funded pools pay out the moment a health behavior is verified — sleep, workouts, preventive care. A Chainlink Confidential AI Attester judges the goal inside a TEE, so raw health data never touches the chain; only the DON-signed verdict is recorded on-chain. World ID gates every pool to one human, one entry, and Arc settles payouts in native USDC. Built across three on-chain integrations in one weekend.
HIPAA-compliant DLP middleware and primary care onboarding agent. Drop it in front of any LLM and patient conversations are automatically scanned and redacted before reaching any model — raw PHI never touches the AI. Patients call a real phone number, get onboarded conversationally, and are triaged straight to the right specialist. No forms, no GP appointment just to get a referral. Built across 3 live components in 6 hours, then shipped post-event as an installable MCP server.
A voice-first simulator for the hardest conversations in medicine — counseling on hyped, under-evidenced compounds like peptides. One tiered evidence registry (Tier 1 trials and openFDA down to Tier 3 anecdote) grounds both an AI patient you interview by voice with an evidence-grounded grader, and a Monte Carlo twin that returns honest, quarter-by-quarter outcome bands — with void and excluded paths so it never overstates thin evidence. Extended post-event into a Tavus CVI video concierge (Maya) that counsels live, cites every claim on screen, and refers to a licensed provider — never prescribing.
Multi-agent due diligence system for startup pitch analysis. Founders upload a pitch deck, type their pitch, or describe their idea and receive a full investment-grade report — competitor landscape, financial stress-testing, and an invest-or-pass verdict from a simulated partner panel. Built on 3 parallel OpenClaw nodes running specialized subagent clusters, powered by Lightning AI’s VPS.
AI wellness coaching app that delivers personalized, real-time video coaching by combining Apple Watch health data, Google Calendar context, and Claude-generated recommendations — delivered via a live Tavus CVI video avatar. iOS Shortcut syncs HRV, sleep, steps, and resting heart rate to a FastAPI backend; a 6:30 AM cron pre-builds context for instant session start. Came back a second weekend after not finishing the first — shipped and demoed.
Clinician-in-the-loop rehab protocol system. Patients chat with an AI coach named Maya; when she detects symptom changes, weekly milestones, or check-ins, a multi-agent LLM pipeline drafts protocol revisions. Every revision goes through a clinician dashboard for approve/reject before flipping live — human oversight on every treatment adjustment. Initial 5-hour demo ran on a GitHub-PR-as-message-bus; productionized post-event by replacing the PR-bus with direct Supabase writes under RLS and upgrading the drafter to a five-agent deterministic pipeline.
Three agentic surfaces on one stack. Drop a GitHub repo and get a research-paper draft for a matched venue (LaTeX + BibTeX) plus an installable Claude Code plugin. Draft purpose-driven outreach (visa, speaking, collab, network) off a Senso brand kit. Track O-1A visa progress against the 8 USCIS criteria with a declared-evidence ledger, AI-streamed petition-quality narratives, and a reportlab PDF dossier for attorney handoff. Built at Datadog HQ in one day; hardened post-event for two real users with passcode auth, per-user data isolation across seven ClickHouse tables, and idempotent migrations. Every LLM call traced through Datadog LLM Observability.