AI agents, health tech, multimodal systems, and everyday tools — things I built because I wanted them to exist.
Healthcare & health tech
Ward.AI
An AI triage system that knows when to call a human doctor
The problem
Most telehealth tools make you fill out a form. Ward.AI actually listens — conducting real-time patient intake and escalating high-risk cases to clinicians before they become emergencies.
What I built
A multi-agent orchestration pipeline with a human-in-the-loop safety layer. The system separates fast-response triage from deeper reasoning to reduce latency, and defers to a human clinician any time it detects uncertainty or elevated risk.
Early cardiac and respiratory screening from a short video and voice clip
The problem
Most people don't get cardiovascular screenings until something goes wrong. EchoCheck explores whether passive physiological signals — things your phone can already capture — could flag risk earlier.
What I built
A multimodal system that extracts heart rate variability and respiratory patterns from facial video (rPPG) and audio inputs. Built feature extraction pipelines for both visual and vocal biomarkers, with structured risk scoring outputs designed for clinical logging.
Stack
PythonComputer VisionrPPGAudio ProcessingGemini API
A voice-first AI copilot that actually controls your Mac
The problem
Most AI assistants answer questions. Jarvis executes — finding files, launching apps, and automating workflows through natural speech, hands-free.
What I built
Designed a "Brain + Brawn" architecture — cloud-based LLM reasoning paired with local Swift executables for fast, reliable system-level task execution. Integrated ElevenLabs for real-time speech synthesis. The split architecture reduced latency by ~40%.
Real-time emergency alert system that calls you when danger is near
The problem
Emergency alerts are slow, generic, and easy to miss. Eleos watches live incident feeds continuously and reaches out to you personally — by phone call — when a hazard crosses into your area.
What I built
An autonomous pipeline that ingests live county emergency feeds via Firecrawl, uses Gemini to parse unstructured incident data, and computes geographic hazard radii to determine user risk. Triggers Twilio voice calls and SMS when severity thresholds are crossed. Built as a monorepo with three decoupled services.
A multi-agent AI system that fights your landlord so you don't have to
The problem
Tenants — especially first-timers — often don't know their rights, can't afford lawyers, and get steamrolled in disputes. Rent-Swarm gives them a system that knows the law and negotiates for them.
What I built
A multi-agent platform with specialized agents for legal reasoning, financial analysis, and negotiation, coordinated through a RAG pipeline. Achieved ~35% improvement in task completion accuracy through decomposed agent workflows.
Upload a product video. Find out if it's worth buying in seconds.
The problem
Secondhand shopping is full of overpriced listings. ThriftSmart removes the guesswork — you show it the item, it checks the market, and it tells you straight: cop or drop.
What I built
A multimodal AI pipeline where Gemini identifies the product from a user-uploaded video, LangChain reasons over live market data, and Selenium scrapes real-time eBay listings for price comparison. A decision engine outputs a clear verdict with a voice explanation via ElevenLabs.
Automated alerts for the cheapest flights to wherever you want to go
The problem
Flight prices shift constantly and checking manually means you always miss the dip. I wanted something that watches for me and sends an alert the moment a deal appears.
What I built
A Python automation tool that searches for the cheapest available flights within a 6-month window using the SERP API. Built with OOP to separate search, filtering, and notification logic cleanly. Integrated SMTP email automation to deliver daily deal alerts.
Gesture-controlled virtual windows using nothing but a standard webcam
The problem
Spatial computing feels like the future, but the hardware is expensive and inaccessible. AirDesk asks: what if you could interact with floating virtual windows using just the webcam you already have?
What I built
A real-time gesture interaction system built in five layers — input, perception, gesture, interaction state, and rendering — each cleanly separated so the system stays extensible. MediaPipe tracks hand landmarks frame-by-frame; a custom gesture engine derives a smoothed cursor and detects pinch signals with hysteresis. Also supports live macOS cursor and window control through Accessibility APIs.