🎙️ AriaPitch Cornell Health AI Hackathon · Pitched at Columbia LionsCage

A web-based voice health screening tool. Records voice samples, extracts acoustic biomarkers, and uses a trained ML ensemble to screen for benign vocal fold lesions.

• 🧠 ML ensemble (SVM + RF + XGBoost) trained on B2AI voice data
• 📊 129 acoustic features extracted via OpenSMILE + Praat
• 🚦 Flag system with actionable recommendations

React Flask scikit-learn XGBoost OpenSMILE

🏥 MedClear · 🏆 Track Prize Winner Columbia Health Hacks · Underserved Demographics Track

A deterministic medical billing verification engine. Programmatically audits hospital bills by cross-referencing line-item charges against Medicare Fee Schedules.

• 🔍 7 independent, composable verification engines
• 📡 Live CMS Medicare Physician Fee Schedule API queries
• 💵 Produces structured, dollar-quantified discrepancy reports

FastAPI Python CMS API Data Parsing

🛡️ RecallGate MongoDB Harness Engineering Hackathon

An on-demand monitor for cyber-capable agents. Reads an action's cheap system footprint and escalates to an expensive LLM monitor only on unfamiliar behavior, then remembers every verdict so monitoring gets cheaper over time.

• 🧩 Two-tier cheap/expensive design with a hard-signal safety override
• 🔎 $vectorSearch memory of incidents in MongoDB Atlas
• 📉 Attack recall 1.00 while expensive-call rate bends downward

Python MongoDB Atlas Vector Search LLM

⚡ Temporal Semantic Caching for Agentic Pipelines Extension of IBM's AssetOpsBench

A temporal query classifier + parallel execution layer that optimizes latency in industrial LLM agent pipelines.

• ⏱️ 3.48× lower median end-to-end latency
• 🚀 296× less tool-discovery overhead
• 🔬 Exposed semantic similarity as an unreliable cache-validity signal

LLM Inference FAISS MCP GCP

Languages

ML / AI

Data Engineering

Frameworks & Tools

Cloud & Databases