Real-time mock interview coach that processes webcam and microphone simultaneously across three streams. Face and voice emotion models are trained from scratch on FER+ and RAVDESS datasets using PyTorch CNNs. Whisper-base handles local transcription, and a custom fusion network combines all three into continuous confidence and engagement scores delivered every 500ms via WebSocket.

Turns any photograph into a clean Monument Valley style isometric illustration. Fine-tuned a Stable Diffusion 1.5 LoRA on ~1,000 synthetically generated isometric images spanning 120 subject categories, trained entirely on Apple Silicon with AdamW, linear warmup, and cosine annealing across 2,000 effective steps. A Canny ControlNet preserves the structure of the source photo while the LoRA restyles it into flat-color geometry. No third-party image APIs are called at any point in the pipeline.

An autonomous LLM assistant that reasons about calendar events, biometrics, and device state to execute multi-step routines without being explicitly asked. Orchestrates 10+ live integrations (Spotify, Google Calendar, Garmin, smart home) via a tool-calling agent loop. Trains local stress and sleep classifiers on wearable data using scikit-learn, with a confidence system that learns from user feedback and adapts in real time.

Fine-tuned DistilBERT on 40K+ labeled Reddit and Twitter posts for 3-class sentiment classification, achieving 73.1% F1 macro on a 10,360-sample held-out test set. Built an analytics layer that computes rolling sentiment averages across configurable time windows, detects anomaly shifts with severity scoring, and surfaces TF-IDF keywords per class with SHAP token-level explanations.

Trained Random Forest, XGBoost, and a PyTorch MLP on 89K Spotify tracks across 7 mood classes. XGBoost reached 99.6% and MLP reached 93.5% F1 macro, illustrating how tree-based and neural models learn threshold boundaries differently. Engineered 13 features from raw audio data and used SHAP TreeExplainer to generate plain-English mood explanations, powering a recommendation engine with cosine similarity ranking and fine-tune sliders.

Educational medical triage system that parses free-text symptom descriptions into structured clinical data using LLM extraction, then scores patient risk with scikit-learn models, reaching 0.92 F1 on risk classification. Ships with a SQLAlchemy schema, FastAPI service layer, a Streamlit dashboard, and a test suite. Built strictly as a demonstration of the modeling pipeline, not for clinical use.

Separates real marketing ROI from confounding signal using double machine learning rather than naive correlation. Estimates both average and individual treatment effects while controlling for age, income, and membership duration, then contrasts causal lift against the correlational read in an interactive dashboard. Containerized with Docker and wired to GitHub Actions CI.

Computer vision and tactical analytics pipeline for broadcast soccer footage. Detects and tracks players, goalkeepers, and the ball using YOLOv8m and ByteTrack, maps pixel positions to real pitch coordinates via homography, and computes formation lines, pressing intensity scores, and Voronoi space control per team.

Deep learning trajectory prediction for multi-player sports footage. Given 50 frames of player tracking data, predicts the next 50 frames using a TemporalCNN and Seq2Seq Transformer. Achieves 0.065 ADE / 0.118 FDE in normalized pitch coordinates, with uncertainty estimation via Monte Carlo dropout.