Things I've Built.
Machine learning system that analyzes supplier transaction data to predict delivery delays, cancellations, and profit volatility. Combines these risks into a composite supplier risk score and visualizes results in an interactive procurement analytics dashboard.
- Engineered supplier-level performance features (late delivery rate, cancellation rate, margin variability) from raw transaction data using Python, Pandas, and SQL.
- Trained XGBoost models to estimate late delivery risk, cancellation risk, and profit margin instability, then combined predictions into a weighted composite supplier risk score.
- Translated model outputs into interpretable supplier risk tiers (Critical / High / Medium / Low) to support procurement prioritization and risk monitoring.
- Built an interactive Streamlit dashboard displaying supplier risk rankings, KPI summaries, department-level risk heatmaps, and risk distribution analytics.
- Analyzed supplier performance patterns to identify high-risk vendors and operational risk drivers across departments.
A full end-to-end data engineering and ML platform that ingests real satellite orbital data from CelesTrak (used by NASA & ESA), engineers collision-risk features from orbital mechanics, trains a risk prediction model, and serves results through an interactive dashboard — mirroring actual Space Traffic Management systems.
- Built a Python ETL pipeline that ingests real orbital element data from the CelesTrak API, parses TLE fields, and loads 9,000+ satellites and 25,000+ debris objects into a normalized SQLite database.
- Computed altitude and velocity from orbital mechanics (mean motion → orbital period → altitude) and engineered collision-risk features from scratch
- Trained an XGBoost classifier to estimate collision risk probability for satellites based on orbital density, debris proximity, and altitude-band characteristics.
- Deployed an interactive Streamlit dashboard visualizing satellite risk scores, orbital congestion patterns, and high-risk altitude zones.
- Analyzed ~35,000 orbital objects and identified high-density altitude bands in Low Earth Orbit where collision risk is concentrated.
A structured technology transfer and commercialization evaluation report for a patented 3D depth-sensing background segmentation system. Assessed market opportunity, competitive landscape, industry applications, and go-to-market strategy across video conferencing, AR/VR, robotics, and surveillance sectors.
- Analyzed the patented computer-vision method (depth + RGB fusion) and identified key technical differentiation enabling lightweight edge deployment compared to neural segmentation models.
- Assessed potential application markets including video conferencing, AR/VR systems, robotics vision pipelines, and intelligent surveillance platforms.
- Evaluated commercialization pathways (IP licensing, strategic partnerships, startup formation) and recommended licensing to computer vision platform vendors as the most viable near-term strategy.
- Conducted competitive benchmarking against depth-sensing platforms (e.g., Kinect lineage) and modern neural segmentation approaches to identify adoption barriers and market risks.
End-to-end reinforcement learning system trained to dynamically price products in a simulated competitive retail market. Calibrated on 26,342 real retail transactions, the PPO agent competes against 5 distinct competitor archetypes across 5 market shock scenarios — outperforming an expert-designed benchmark in all conditions with profit lifts of +8% to +79%.
- Cleaned and processed 26,342 real retail transactions to fit a calibrated OLS demand model (R²=0.516), producing 13 publication-ready validation figures.
- Built a custom OpenAI Gymnasium environment with a composable Shock Engine that injects 5 market disruption types (inflation, recession, demand surge, stagflation) by modifying structural demand parameters — not arbitrary multipliers.
- Trained a PPO reinforcement learning agent over 600,000 simulation steps across 5 random seeds; agent wins 5/5 market scenarios and captures 46.8% of the revenue pool vs. a 33.3% equal-share benchmark.
- Introduced a price-stability penalty (λ=2,500) to the reward function that reduces price volatility by 90.8% while simultaneously improving profit by 2.1% — addressing a key barrier for real-world e-commerce deployment.
- Benchmarked against 5 competitor archetypes (FixedSeller, Undercutter, Premium, Chaotic, Reactive) across 25 head-to-head matchups; agent beats 4/5 archetypes with structured analysis of the Undercutter boundary condition.
Currently building. Follow on GitHub for updates.