A comprehensive collection of production-grade machine learning solutions, agentic systems, research pipelines, and data architectures I have developed.
Final Reinforcement Learning course project where we implemented PPO from scratch, validated it on CartPole-v1 (mean return 500), and trained it on racetrack-v0 (highway-env). It includes custom PPO vs SB3 comparisons and demo rollouts showing stable lane-following with smoother steering.
End-to-end AI platform for CV understanding, job opportunity matching, and recruiter-side candidate discovery. Powered by a custom fine-tuned Qwen model on Google Cloud, FastAPI, Next.js, and Supabase, featuring hybrid search and fallback ranking under high traffic.
Reproducible inventory decision prototype built on the public Walmart M5 dataset. It combines probabilistic LightGBM forecasts, (s,S) and MILP replenishment policies, and 200 SimPy replications. In the documented simulation, MILP reached a 98.10% fill rate and reduced expected cost by 36.66% versus the baseline under synthetic costs and lead times.
Machine Learning project to estimate used SUV prices from 22,377 listings. I trained and tuned four regression models through Azure Machine Learning jobs and datasets on dedicated compute clusters; Random Forest achieved the best test metrics (RMSE 7656.09, MAE 4784.67), with mileage and vehicle year analyzed as key price drivers.
Kaggle Playground S6E5 binary classification project predicting whether an F1 car pits on the next lap. The pipeline combines LightGBM, XGBoost, CatBoost, and a PyTorch MLP with 5-fold cross-validation, out-of-fold target encoding, seed averaging, rank blending, and about 0.95 ROC-AUC.
Uplift/incrementality models on the Hillstrom dataset (64k customers, randomized treatment) to decide who to target in a marketing campaign - estimating the causal effect of an email, not just who buys. Compares S-Learner, T-Learner, and Class Transformation with Qini/AUUC, hyperparameter tuning, and robust multi-split evaluation that shows the methods are tied within noise.
Interactive Tableau dashboard for analyzing AI infrastructure costs, allocated budgets, business value, and ROI. It revisits concepts introduced in my Database course and extends them into a decision-oriented BI use case with a synthetic 16,200-row dataset.
Final Artificial Vision project developed with a teammate, comparing YOLOv11 detectors across RGB, thermal, GST, HST, and late-fusion strategies. We exported the trained model from .pt through ONNX to a TensorRT engine and ran FP16 inference on an NVIDIA GeForce RTX 3090 to reduce memory requirements and accelerate inference. On new, domain-similar images, the fine-tuned model was qualitatively more robust than the baseline on unclear cases; evaluation also included confusion matrices and PR/F1 curves.
Final Autonomous Robotics project built in ROS 2 with Python and C++. We developed a custom SLAM pipeline with particle filtering, occupancy-grid log-odds, and LiDAR ray tracing, plus an AMCL-style localization and navigation scaffold using TF2 and Nav2. After validating the system in Gazebo and RViz2, we transferred it to a physical TurtleBot3 Burger and tested mapping, localization, and navigation on a university laboratory track with the robot's sensors.
Computer vision pipeline to reconstruct panoramas from three images, implemented end-to-end in Jupyter. It includes SIFT feature extraction, ANMS keypoint selection, robust matching with Lowe ratio and bidirectional checks, homography estimation with DLT + RANSAC, and weighted blending for smooth final reconstruction.