Matthew Ohanian

Full-Stack AI Engineer - California University of Science and Medicine

• Building source-grounded healthcare RAG backend and privacy-first cross-platform mobile app for MS caregiver support in collaboration with physicians and medical students at California University of Science and Medicine
• Designed 5-stage RAG pipeline (guard rails, HyDE, hybrid vector+BM25 retrieval, FlashRank reranking, grounded generation) passing a 90-example source-anchored eval with 1.000 Recall@5, 0.851 MRR, 0.896 faithfulness, and 1.000 no-answer success
• Engineered crisis detection, prompt injection defense, and scope detection for healthcare-sensitive queries with self-auditing citation enforcement
• Applied lost-in-middle reordering, near-duplicate removal, and source diversity capping to optimize LLM context quality
• Built React Native + Expo app (iOS, Android, Web) with baseline MSSS-5 social support, daily MHI-5-style mood/support check-ins, weekly GAD-7/PHQ-9 gated screeners, PHQ-9 self-harm crisis routing, and privacy-first AsyncStorage architecture

Data Scientist - Qventus

• Developed surgical case prediction models using ensemble methods with XGBoost and LightGBM on healthcare snapshot data with Python (pandas, numpy, PyTorch, scikit-learn, imbalanced-learn, SMOTE)
• Built ensemble models with SMOTE for class balancing on AWS SageMaker, reducing false negatives by 35%
• Built real-time inference MLOps pipelines on AWS SageMaker with latency optimization and automated monitoring
• Engineered 50+ features using Snowflake SQL for production ML systems with automated model monitoring
• Applied transformer-based embeddings (BERT/Hugging Face) to surgical procedure text for semantic similarity and case clustering
• Implemented automated model retraining pipelines with MLflow experiment tracking and model versioning
• Partnered with clinical teams to translate requirements into ML solutions, presenting weekly model evaluations to stakeholders

Data Scientist - Raydiant

• Architected ETL pipelines using AWS Glue (Crawlers, Jobs), S3, Lambda, and Athena to process 3M+ customer records with automated data quality checks, reducing pipeline errors by 45%
• Developed churn prediction model achieving 85% AUC-ROC with SHAP analysis identifying key retention features, enabling targeted customer interventions
• Built Tableau and Salesforce dashboards for leadership, enabling self-serve KPI monitoring and faster decision-making on retention initiatives

Data Scientist - Axaitech

• Built 11-class cancer classification model using deep learning, neural networks, and ensemble methods, achieving 95% accuracy while reducing false positives by 40% vs. baseline
• Applied Random Forest feature selection to reduce gene expressions from 58,000 to 16 critical biomarkers without accuracy loss
• Implemented ensemble methods (Random Forests) with cross-validation and GridSearchCV for robust model validation
• Developed automated data preprocessing pipeline using Python and scikit-learn, improving model training efficiency by 60%

Data Scientist Consultant - DataRes

• Built NLP-powered dashboard using TF-IDF and sentiment analysis for 10K+ real estate listings, improving client engagement by 30% through data-driven content recommendations
• Developed OCR pipeline with PyTesseract achieving 92% accuracy on 5K+ wine labels
• Created interactive visualizations using Plotly and React to communicate insights to stakeholders

Mathematics Instructor - Matt's Tutoring Services

• Tutored 25+ students in Calculus, Linear Algebra, Statistics, and Data Science, improving grades by 20-30% through personalized instruction and statistical modeling concepts
• Managed and delivered tutoring sessions both in-person and virtually via Zoom, maintaining high student engagement
• Coordinated schedules around students' commitments including sports and extracurricular activities
• Developed customized learning plans and practice materials tailored to each student's learning style and goals

Mathematics Instructor - Mathnasium Learning Center

• Provided personalized mathematics instruction to 50+ students from elementary through high school, specializing in advanced concepts for middle and high school learners
• Designed customized learning plans based on comprehensive student assessments, resulting in measurable improvement for 85% of students within three months
• Implemented positive reinforcement techniques to boost student confidence and motivation, leading to increased engagement and academic performance
• Maintained regular communication with parents to discuss progress and adjust learning strategies, resulting in high satisfaction ratings

InDaCapitol - Congressional Stock Trade Predictor

• Built end-to-end ML system that scrapes, parses, enriches, and analyzes U.S. Congressional stock trades to estimate buy outperformance and sell timing alpha versus SPY
• Engineered daily ETL pipeline integrating 6+ data sources (House/Senate filings, yfinance, Finnhub, Congress.gov, USAspending) processing 14K+ transactions with dual PDF parsing strategy
• Built 3-tier ticker extraction system: SEC CIK mapper, fuzzy matching (rapidfuzz), and Groq LLM fallback for ambiguous company names
• Designed segmented feature sets (17 buy features, 19 sell features) covering committee-sector overlap, politician history, sell-history signals, filing delay, and market context
• Trained June 10, 2026 active LightGBM classifiers: buy AUC 0.695 / F1 0.640 / accuracy 64.0%, sell AUC 0.697 / F1 0.689 / accuracy 63.9%
• Trained buy/sell LightGBM regressors for 30-day SPY alpha with MAE 0.0813 / 0.0774, RMSE 0.1117 / 0.1114, and R² 0.177 / 0.185
• Current monitoring: 16,687 feature vectors, 16,638 predictions, 22 registered models, 99.9% feature coverage, 99.6% prediction coverage, and latest transaction on June 4, 2026
• Realized health monitor shows 73.3% classifier accuracy over 16,144 samples, 0.0566 regression MAE, and stable feature drift
• Shipped production Next.js dashboard on Vercel with Supabase PostgreSQL, prediction badges, model registry-backed batch predictions, FastAPI health/models/predictions endpoints, and monitoring for promotion decisions
• Technologies: Python, LightGBM, scikit-learn, SQLAlchemy, FastAPI, pdfplumber, Gemini Vision OCR, Next.js 16, React 19, Tailwind CSS, shadcn/ui, Recharts, Supabase, GitHub Actions, Vercel

JobLens - Privacy-First Job Email Classifier

• Built Chrome extension that classifies job emails (applications, rejections, interviews, offers) using on-device ML inference - no data leaves the browser
• Trained on 28,500 emails (20.5k real + 8k synthetic via DeepSeek/Ollama) with 3-phase curriculum learning to achieve 90%+ accuracy on Gmail's ~200-character truncated snippets
• Finetuned BAAI/bge-small-en-v1.5 with custom classification head; INT8 quantized from 128MB to 34MB (73% reduction)
• Optimized for browser-edge deployment via Transformers.js + ONNX WebAssembly; published on Chrome Web Store

EduAgent - AI-Powered Educational Video Generator

• Built a multi-agent AI architecture using CrewAI and Anthropic Claude for intelligent content analysis and lesson planning
• Integrated Manim (3Blue1Brown's animation engine) for creating professional mathematical visualizations
• Implemented natural text-to-speech narration with synchronized audio-visual timing
• Developed a clean web interface with Gradio for easy content upload and real-time video generation
• Technologies: Python, Anthropic Claude API, CrewAI, Manim, MoviePy, LMNT TTS, Gradio, OpenCV, PyPDF2
• Impact: Democratizes high-quality educational content creation by reducing video production time from hours to minutes while maintaining professional quality standards

Hierarchical Vision Transformers for NIH Chest X-Ray Classification

• Built a two-stage Vision Transformer pipeline for NIH Chest X-ray classification, using a binary finding detector to gate a 14-label disease classifier
• Designed the hierarchy to mirror radiology workflow while reducing class-imbalance pressure from the dominant "No Finding" class
• Implemented reusable preprocessing and data-loading utilities with 224x224 resizing, ImageNet normalization, balanced sampling, caching, and multiprocessing
• Evaluated the hierarchical model at 0.7563 mean disease AUC, with stronger performance on Cardiomegaly (0.8861), Consolidation (0.8793), and Pneumonia (0.8535)
• Documented limitations transparently, including binary-gate calibration concerns from a 1.0 accuracy / NaN AUC result and downstream error propagation risk

AI Poker Arena - LLM Battleground

• Built autonomous AI battleground where 6 frontier LLMs (Claude, GPT-5, Gemini, Grok, Mistral, DeepSeek) compete in Texas Hold'em poker
• Ran 15,000 hands across 15 round-robin matchups, discovering GPT-5 dominates (+143 BB/100) despite "calling station" style
• Implemented real-time opponent profiling system tracking VPIP, PFR, aggression factor, and adaptive strategy classification
• Developed multi-provider LLM integration layer with LangChain for structured output parsing across different API formats
• Technologies: Python, LangChain, Pydantic, SQLite, Plotly, Pandas, Weights & Biases

Advanced ML Competition: Childhood Growth Prediction (PyTorch, MLflow, Funk SVD)

• Achieved 5th/90 in a competitive ML challenge using matrix factorization and PyTorch neural networks for childhood growth prediction
• Tracked 50+ experiments with MLflow for model versioning, tracking, and performance monitoring across iterations

E-commerce Cart Abandonment Prediction (PySpark, LightGBM, A/B Testing)

• Engineered 45+ behavioral features from 2M+ web logs achieving 85.7% accuracy with LightGBM for customer retention
• Designed A/B testing framework with statistical power analysis, projecting $500k annual revenue increase with 95% confidence through targeted interventions

Winosaurus OCR Model (UCLA DataRes)

• Built an OCR model to read wine labels for Winosaurus
• Processed images to improve OCR accuracy and tuned the model to optimize performance
• Created data visualizations to communicate results to stakeholders

HomeDescription Dashboard (UCLA DataRes)

• Built a dashboard to provide real-time feedback on home descriptions
• Analyzed 10,000+ listings to identify key factors for content improvement, resulting in 30% higher engagement rates
• Developed data visualizations to communicate insights about effective home descriptions

PAYGo Solar Technology Project (iXperience)

• Built a machine learning model to predict customers' next month's and six-month payments based on previous payment history
• Performed PCA to reduce data dimensionality and improve model performance
• Trained a linear regression model to predict customer payments with evaluation on held-out test data

Happiness Score Project (UCLA Math-156-Final-Project)

• Implemented and compared multiple statistical modeling approaches to analyze happiness scores, including multivariate regression, KNN, and Principal Component Analysis (PCA)
• Developed custom mean squares algorithm implementation and compared results against standard library functions
• Collaborated with a team to analyze, clean, and model happiness score data across different countries and factors
• Calculated Euclidean distances for K-nearest neighbors classification focused on social support as a response variable
• Created comprehensive mathematical derivations in LaTeX for model documentation and analysis

Master's of Science in Data Science (MSDS)

• GPA: 3.9/4.0
• Coursework: Linear Regression, Relational Databases, Computation in Python, Statistics, Data Structures and Algorithms, Distributed Computing, Machine Learning, Deep Learning, Ethics for Data Science and AI, Generative AI, Experimental Design, Distributed Data Systems, Data Engineering

Bachelor of Science in Data Theory

• GPA: 3.7/4.0
• Dean's Honors List: Fall 2021, Winter 2022, Winter 2023
• Roles: Data Scientist at UCLA DataRes and Vice President at UCLA's Club Sampling Club
• Coursework: Machine Learning, Math Modeling, Python and Technologies for Data Science, Data Analysis and Regression, Statistical Programming with R, Algorithms, Computation and Optimization, SQL

Data Science Certificate

6-week educational experience studying and interning in Data Science, focusing on Python programming, data preprocessing and cleaning methods, data analysis and feature engineering, data visualization, and machine learning techniques

Associate for Science in Transfer: Mathematics

• GPA: 3.9/4.0
• Dean's Honor Roll: Fall 2019, Spring 2020, Fall 2020, Spring 2021
• Parker Mathematics Scholarship Recipient

Community Service

• Lower Nob Hill Weekly Cleanup, San Francisco, CA (January 2024-July 2025): Participated in weekly Sunday community cleanup events to maintain and beautify the neighborhood.
• Project Paper Bag, Los Angeles, CA (2019-2023): Prepared and delivered meals for the homeless
• Olive Crest, Riverside, CA (2021-2022): Assisted in organizing fundraisers to support children in foster care
• ACEing Autism, Burbank, CA (2017-2018): Helped teach children with autism how to play tennis