Sandro Amaglobeli

Trading Operations Risk Intern, Software Engineering

Built automation and monitoring tools for PSEG’s trading operations to reduce manual reconciliation risk and improve reliability of energy-auction workflows. Developed Python ETL pipelines with validation/retries, migrated legacy VBA/VBS/Batch jobs into logged Python/SQL services, and building a FastAPI dashboard for real-time visibility and alerting during data outages

Blind-IGT: Jointly Decoding Rewards and Rationality in Entropy-Regularized Competitive Games

Co-authored a framework resolving a fundamental identifiability problem in inverse game theory—jointly recovering agent rewards and rationality when neither is observed. Designed a Normalized Least Squares estimator achieving optimal O(N^{-1/2}) convergence rates. Generalized to entropy-regularized Markov games with proven robustness under unknown dynamics.

Modeling human observer performance with neural network observers in a forced localization task using undersampled MRI images

Built FLNet (EfficientNet-B1 variant) for tumor localization under MRI undersampling artifacts, trained on 8,240 images spanning multiple artifact regimes. Developed a vision-inspired variant using Gabor-based preprocessing to better match human perception. Achieved 84% accuracy (vs. humans 82% and SDOG4 68%) and improved generalization to extreme / out-of-distribution artifact conditions.

Reachability-Constrained RRT* for Car-Trailer Parking

Built RC-RRT* parking planner for car-trailer nonholonomic dynamics using reachability-guided expansion and tube-verified edges. Verified hitch safety via 1D zonotope interval propagation with conservative linearization for inflated collision geometry, maintaining |γ| ≤ 70° under sensor uncertainties. Achieved 100% planning success with up to 49× speedup vs RRT* and Hybrid A*.

Multi-View Generalizable and Animatable Gaussian Head Avatar

Extended GAGAvatar to multi-view with visibility-aware voxel fusion and dual-layer Gaussian lifting to improve surface completeness. Fine-tuned 2.41% of parameters via LoRA PEFT, achieving LPIPS 0.623 and CosSim 0.721. Built interactive digital human pipeline integrating NVIDIA Audio2Face with 52-D blendshape-to-FLAME mapping and LLM-powered conversation.

Traffic Flow Anomaly Detection

Implemented real-time changepoint detection algorithms (CUSUM, Page-Hinkley) and offline segmentation (PELT) to identify significant shifts in traffic duration data. Analyzed trade-offs between detection delay and false alarm rates by tuning tolerance thresholds and forgetting factors.

• Implemented Page-Hinkley and CUSUM algorithms from scratch for online anomaly detection
• utilized PELT (Pruned Exact Linear Time) for optimal offline trajectory segmentation
• Analyzed sensitivity vs. specificity trade-offs for robust event detection in noisy time-series data

3D Gaussian Splatting for Scene Reconstruction

Built real-time photorealistic 3D scene reconstruction pipeline from 2D images. Architected CUDA-accelerated Gaussian splatting with a lightweight cross-platform viewer enabling interactive exploration of reconstructed spaces.

Traverse - AI-Powered Travel Planning Platform

Built a travel planning web app that generates city itineraries and visualizes attractions and restaurants on an interactive map. Integrated Airbnb listings with data insights (neighborhood/host analytics, violin plots, heatmaps) and regression-based price modeling to highlight high-value stays. Won 1st Place at Hofstra Hacknology 2024.

AI Legal Assistant Chatbot for NY Small Claims Court

Built AI chatbot guiding users through NY Small Claims Court filings with 53-language support. Designed ML classifier achieving 90% accuracy across 100+ disability subcategories using Python, Rasa, and JavaScript. Led second phase implementing form autofill based on guided user interviews.

RealTick - Real-Time Stock Analysis Dashboard

Full-stack stock analysis dashboard delivering real-time market data across 5,683+ tickers. Includes technical indicators, interactive charting, stock comparison, watchlists, and a curated news feed. Flask REST API aggregates Yahoo Finance data, and a Vite/React frontend provides a responsive, professional dark-themed UI.

• Real-Time Data - Live stock price updates with minute-level data
• Technical Analysis - Ichimoku Kinko Hyo, Bollinger Bands, RSI, MACD, and moving averages
• Interactive Charts - Candlestick, area, line, and OHLC charts with zoom and timeline controls
• Stock Comparison - Compare multiple stocks side-by-side for market analysis
• Financial News - Curated news feed with publisher info sorted by publish date
• Watchlist - Track and manage your favorite stocks
• Similar Stocks - Browse stocks in the same sector and market cap range
• Trending Stocks - View currently trending and popular stocks
• REST API - Full-featured backend with endpoints for historical and real-time data

GraphX - Balance of Payments Interactive Visualization

Built interactive web portal for Georgia's central bank visualizing national Balance of Payments data to support economic analysis and policy decisions.