Built a full Retrieval-Augmented Generation (RAG) chatbot embedded in this portfolio. Visitors can ask natural language questions about my background and get accurate, streamed answers. Uses Voyage AI for semantic embeddings, Supabase pgvector for similarity search, and Claude Haiku for response generation.

Autonomous trading system monitoring 4,100+ live Polymarket prediction markets, ingesting 120 news articles per cycle from Polygon.io and 6 RSS feeds (Reuters, AP, CNBC, CoinDesk). Features a swarm AI engine ("MiroFish") orchestrating 10 GPT-4o-mini analyst personas for crowd-consensus probability estimates, triggering trades when AI consensus diverges >15 pp from live prices. Kelly Criterion position sizing with confidence scaling, a 5-layer modular pipeline with 130 unit tests, a CircuitBreaker that auto-halts on 3 consecutive losses or 10% daily drawdown, and a real-time "Kinetic Terminal" dashboard (FastAPI + React/Vite + Recharts) with live signals and P&L analytics.

Full-stack data analytics platform serving 20,000+ UTD students with grade distributions, professor ratings, and course insights. Led two major UI redesigns, increasing engagement by 500% and improving load speed by 30%.

Scalable student note-sharing platform built for UTD. Leading 30+ engineers and designers, managing sprints, onboarding, and code reviews. Architected the platform using Next.js, MUI, Drizzle ORM, and NextAuth.

Built an automated productivity & stress companion for product managers. Extracts action items from meeting notes, auto-approves high-confidence backlog stories into Notion, and provides wellness insights via a live dashboard. Integrates Google Calendar, Notion API, Google Gemini, and ElevenLabs for voice.

Developed assistive technology translating images into descriptive audio for visually impaired users. Integrated Google Cloud Vision and Azure Computer Vision APIs for real-time image recognition. Utilized gTTS for speech synthesis and OpenAI for intelligent caption generation.

Built a deep learning sentiment analysis model classifying online comment toxicity across multiple labeled categories. Applied Recurrent Neural Networks (RNNs) to capture linguistic context. Processed and augmented datasets to improve generalization and model performance.