β FEATURED // RAG SYSTEM
DocuQuery
Production Project
RAG Document Intelligence System - Hybrid vector search with recursive splitting,
Cohere reranking, and async SSE streaming for sub-2s query latency.
+30% Answer Precision
<2s Latency
Hybrid Vector Search
FastAPI Β· OpenAI Β· Pinecone Β· LangChain Β· React Β· Docker Β· AWS
β FEATURED // AGENTIC PLATFORM
Share Cycle
Production Project
Agentic Support & Orchestration Platform - Built concept-to-production in 24
hours using an AI-native development workflow with full auth and data layer.
24hr Build-to-Prod
AI-Native Workflow
Full MVP + Auth
Django REST Β· React Native Β· PostgreSQL Β· OpenAI Β· Auth0 Β· Redis
π FEATURED // F1 PERSONAL
F1 Race Prediction Model
Personal Project
ML model predicting top-5 race finishers using lap delta, weather conditions, and
historical race data. Ensemble approach with gradient boosting.
85% Accuracy
Top-5 Prediction
Lap Delta + Weather
Random Forest Β· Gradient Boosting Β· Python Β· Pandas Β· Scikit-learn
// AI/ML - COMPUTER VISION + LLM
Mood Sync
Dec 2025 β Jan 2026
- Real-time facial emotion recognition pipeline - 95%
precision across 6 emotional classes
- Pinecone vector search with 0.9s average recommendation
latency
- LLM reasoning layer - >85% user-rated relevance in pilot testing
- 30% CPU latency reduction via multithreading + frame skipping; 22β24 FPS live inference
Python Β· Pinecone Β· LLM Β· OpenCV Β· Multithreading
// AI/ML - GENERATIVE VISION
VAE for Emoji Generation
Jun β Jul 2025 Β· University at Buffalo
- Convolutional VAE in PyTorch on 12,000+ emoji images - 89%
reconstruction fidelity (SSIM)
- Reparameterization trick enabling smooth latent interpolation across emotional dimensions
- 27% reconstruction loss reduction via KL annealing; Ξ²-VAE improved separability by 22%
PyTorch Β· VAE Β· Ξ²-VAE Β· Convolutional Networks
// AI/ML - CONTRASTIVE LEARNING
Image-Caption Retrieval (SimCLR)
Mar β May 2025 Β· University at Buffalo
- SimCLR retrieval engine over 31,000 image-caption pairs -
68% Top-5 accuracy
- PCA dimensionality reduction 512β64 retaining 93% variance; t-SNE cluster visualization
- Streamlit demo with 1.2s average query latency
SimCLR Β· PyTorch Β· PCA Β· t-SNE Β· Streamlit
// AI/ML - NLP SUMMARIZATION
Congressional Bill Summarization (BART)
Apr 2025 Β· University at Buffalo
- Fine-tuned BART on 22,000+ congressional docs - ROUGE-L +11%
over baseline
- 18% hallucination reduction via preprocessing; ROUGE-2: 18.6 / ROUGE-L: 41.2
- Dynamic truncation + batching for docs exceeding 1,024-token limit
BART Β· HuggingFace Β· NLP Β· PyTorch
// AI/ML - NLP MULTI-DOC
Multi-News Summarization (BART)
Apr 2025 Β· University at Buffalo
- Fine-tuned facebook/bart-base on Multi-News - ROUGE-L 38.7
- 21% inference time reduction via beam-width tuning; Gradio app with 2.6s response
- Deployed reproducible pipeline via Hugging Face Hub
BART Β· Gradio Β· HuggingFace Hub Β· NLP
// AI/ML - TIME SERIES
Air Quality Prediction (LSTM)
Mar 2025 Β· University at Buffalo
- LSTM forecasting hourly pollutants - RMSE β19% over linear
regression
- Lag feature engineering +14% stability; RΒ² improved 0.61 β 0.79
- LSTM vs GRU comparison: 12% faster convergence observed in GRU
LSTM Β· GRU Β· PyTorch Β· Time Series Β· Pandas
// AI/ML - DEEP LEARNING
VGG & ResNet from Scratch
Feb 2025 Β· University at Buffalo
- VGG: 87% test accuracy Β· ResNet: 90% test accuracy - no pretrained weights
- 29% faster convergence via batch norm + LR scheduling; 18% overfitting reduction
- Grad-CAM validation: 88% of correct samples activated on correct regions
PyTorch Β· VGG Β· ResNet Β· Grad-CAM Β· CNN
// AI/ML - REINFORCEMENT LEARNING
RL Vacuum Cleaner Bot
Nov 2024 Β· University at Buffalo
- 6Γ6 grid-world with stochastic transitions - SARSA 24% faster
convergence vs Q-Learning
- Double Q-Learning reduced overestimation bias by 17%
- Stable optimal policy learned within 450 training episodes
SARSA Β· Q-Learning Β· Double Q-Learning Β· Python
// AI/ML - COMPUTER VISION
EMNIST Character Recognition (CNN)
Oct 2024 Β· University at Buffalo
- Custom CNN - 92.4% test accuracy on EMNIST Balanced
- +11% validation accuracy via augmentation + dropout; training loss β33% via LR scheduling
- ROC + confusion matrix analysis identified top 3 confusion clusters for architecture tuning
PyTorch Β· CNN Β· EMNIST Β· Data Augmentation
// DATA - REAL ESTATE ML
Real Estate Sales Prediction
Feb 2025 Β· University at Buffalo
- Feature engineering on 1.09M+ transactions - 31% missing
data reduction
- Gradient Boosting + XGBoost: RΒ² 0.86, RMSE β17% vs baseline
- Bayesian hyperparameter optimization +9%; Tableau dashboards across 150+ neighborhoods
XGBoost Β· Gradient Boosting Β· Bayesian Opt Β· Tableau Β· Python
// DATA - INFRASTRUCTURE
Environmental Data Warehouse
Mar β Apr 2025 Β· University at Buffalo
- Normalized schema: 10 relational tables for structured environmental analytics
- 34% query performance improvement via indexing + EXPLAIN ANALYZE; stored procs β42% repetition
- Join-heavy query optimization: 2.3s β 1.4s execution time
PostgreSQL Β· SQL Optimization Β· Data Warehousing Β· Indexing
// DATA - PRICING ANALYTICS
Airbnb Price Prediction
Sep β Nov 2024 Β· University at Buffalo
- EDA on 48,000+ Airbnb listings - identified top 5
price-driving features
- Random Forest: +16% accuracy over linear baseline; MSE β21% via feature engineering
- Cross-validated RΒ² of 0.77 across folds
Random Forest Β· Feature Engineering Β· EDA Β· Python Β· Scikit-learn