AI/ML & Full-Stack Engineer building AI-native systems.

Python · FastAPI · React/TypeScript · Whisper · RAG · Ollama

• Records lectures and transcribes them locally on-GPU with Whisper — fully offline via Ollama, or routed to Gemini/Claude
• RAG-backed tutor answers questions with citations back to the exact minute of the recording
• FastAPI backend with a React/TypeScript frontend, built local-first for privacy

FastAPI · SBERT · Gemini · Docker · JavaScript

• Résumé ↔ job-posting matcher using a 3-agent pipeline: extraction, semantic matching, and LLM-generated recommendations
• Weighted scoring across skills, responsibilities, title match, and semantic similarity
• 3× faster inference through caching and async pipelines, surfaced in a responsive JS UI

React · TypeScript · Multilingual UX

• Multilingual (15 languages) web app helping undocumented, uninsured, and low-income patients navigate U.S. healthcare without an account
• Surfaces charity-care programs and sliding-scale clinics, explains medical bills, and helps with applications
• Built and shipped a live, working product end-to-end

CatBoost · LightGBM · Quantile Regression · Calibration — FitchGroup Codeathon 2025

• Achieved R² = 0.65 (Scope 1) and 0.48 (Scope 2) with leakage-safe target encoding and quantile ensembles
• Isotonic calibration on top of the ensemble for well-calibrated uncertainty estimates
• Modeled $174M risk reduction for portfolio-level emissions estimation

MCP · LLM Explanations · Break Through Tech AI Studio

• Recommendation engine matching newcomers to 3–5 neighborhoods from budget, household, and lifestyle inputs
• MCP-backed LLM explains matches in plain language without inventing listings
• Built as a Break Through Tech AI Studio challenge, advised by Flow Miami

Chrome Extension (MV3) · PDF.js · Gemini API

• MV3 Chrome extension that finds and highlights Scope 1 GHG emissions values inside sustainability report PDFs
• PDF.js-based lazy-rendered viewer with in-context highlighting
• Gemini API generates a copyable summary for reporting or pitching

Python · Decision Trees · Feature Engineering — UIC CS218

• Classifier predicting whether a loan will be charged off, trained on historical Lending Club application data
• Feature engineering and encoding across categorical and numeric application fields
• Improved charge-off detection over the ~50% baseline rate through model tuning

Python · FastAPI · Streamlit · WebSockets

• Local-network telemetry pipeline ingesting sensor data from multiple NVIDIA Jetson devices into a central FastAPI server
• Per-device authentication, structured JSON payloads, and per-day CSV export by device and date
• Optional WebSocket streaming plus a Streamlit dashboard for live visualization