I'm a graduate from the University of Southern California (MS Applied Data Science, GPA: 3.73), graduated May 2026. I build production-ready AI/ML systems — from RAG pipelines and multimodal recommenders to privacy-first LLM platforms — with deployments at scale on AWS and Spark.

My work spans Retrieval-Augmented Generation (RAG), Multimodal AI, NLP, Computer Vision, and scalable ML deployment. I'm passionate about creating secure, high-impact AI solutions that automate workflows and enhance real-world decision-making.

• ✦ Origin Weekend: IMPACT S26 – First Place — USC startup launch sprint organized with Google & TIE Hub – USC Viterbi
• ✦ Teaching Assistant — DSCI 551 (Graduate) & DSCI 351 (Undergraduate), Foundations of Data Management
• ✦ Advanced Deep Learning Research @ NUS — Conducted privacy-preserving facial analysis research; improved model accuracy by 10% and reduced TensorFlow inference latency by 40%
• ✦ AWS Certified Cloud Practitioner

• Built an internal Slack-based assistant for Q&A, summaries, and insights using a RAG architecture with the OpenAI/Gemini API, incremental data sync, a Vue frontend, and PII-redaction safeguards — reducing manual information requests and delivering answers within seconds
• Led CRM integration planning (API research, webhook architecture, field mapping) and designed a staged sync workflow that validates payloads before production, improving data integrity and cutting sync errors
• Built national compliance datasets with tiered state coverage, machine-readable exports, late-fee and closure reference indexes, and contributed to a data reconciliation pipeline across internal and third-party systems, enabling faster compliance reporting and reducing manual data-matching effort

• Built RAG system with OpenAI Function Calling + Firebase handling 500+ voice notes, delivering actionable pet health analytics across 3 cross-team dashboards to inform product decisions
• Developed a multimodal AI pipeline processing 200+ audio files with real-time speech transcription, PDF analysis, and summarization using TensorFlow and Transformers, reducing manual document review time by 60%
• Created a visualization engine in Scala supporting 12+ chart types with Redis caching, cutting data representation render time by 45% and enabling real-time analytics for 100+ daily queries

• Built an AI-powered PO Automation conversational AI to analyze large purchase orders and enable multi-document querying, resulting in streamlined processing and improved data accessibility
• Designed and optimized prompt sets with GPT-4 and custom evaluation metrics to test system accuracy, relevance, and robustness in production, resulting in higher query precision and more reliable responses
• Contributed to scalable deployment workflows, improving inference performance and reliability for enterprise use, which led to increased system efficiency and reduced downtime

Fine-tuned Gemma 4 E4B (QLoRA via Unsloth, final loss: 0.018) on WHO/SPHERE/IMCI emergency protocols for frontline triage. Supports 6 languages (English, Swahili, Hindi, French, Arabic, Hausa) with auto-detection, voice input, photo triage (GPT-4o vision), and spoken audio guidance. Offline-capable React Native app + Next.js web app. Published model on HuggingFace.

Privacy-first AI platform with real-time streaming chat, routing across 5+ local Ollama models (Llama 3.2, DeepSeek Coder, Phi3), supporting 10+ file types with semantic chunking. Includes Google OAuth, Redis session management, rate limiting, and GDPR/HIPAA compliance features.

AI-powered team workspace built with React + Gemini. Chat, sprint management, meeting recaps, and departmental announcements — all in one place.

Camera-based waste classification app powered by Amazon Nova on AWS Bedrock, providing real-time bin recommendations (Waste / Recycling / Compost). Features smart frame gating to minimize Bedrock calls, confidence-based agent decision modes, feedback capture, and a SQLite analytics dashboard. Deployed on EC2 with FastAPI, nginx, and Let's Encrypt. Built for the Amazon Nova AI Hackathon.

Spark RDD + XGBoost recommender with 40+ engineered features for cold-start and behavioral modeling. Achieved RMSE of 0.9734 on test set using confidence-weighted biases and GLM techniques.

LLM-powered database visualization tool converting natural language to SQL; supports SQLite, MySQL, PostgreSQL with Matplotlib-based bar, line, and scatter plot generation.

• RAG & LLMs: OpenAI Function Calling, Gemini API, Claude (Bedrock), Prompt Engineering, Foundation Models, QLoRA Fine-tuning
• Deep Learning: Transformers, CNNs, RNN/LSTM, Vision Models, Multimodal Pretraining
• NLP: NER, Sentiment Analysis, Text-to-SQL, Speech Transcription
• ML Engineering: Model Deployment, Quantization, Optimization, Statistical Modeling, Data Mining

• Languages: Python, TypeScript, SQL, Java, C, Scala, Rust
• Libraries: PyTorch, TensorFlow, Scikit-learn, OpenCV, Pandas, Unsloth, TRL
• Web: Next.js, FastAPI, Flask, React.js, React Native, Vue.js
• DBs: Aurora PostgreSQL, MySQL, MongoDB, Firebase, Redis, SQLite, Upstash
• Cloud & DevOps: AWS (CCP Certified), Bedrock, EC2, Lambda, S3, SQS, SNS, EventBridge, Databricks, Spark RDD, Hadoop, Vercel, Clerk, Sentry, Git

• Brain Stroke Detection using ML Models – IEEE Xplore, Feb 2023
  Read Paper