Bazal Khan

Machine learning for messy operational data. Farmingdale, NY · Graduating December 2026.

RTO risk + category classifier Return-to-origin scoring and hierarchical product categorization. PostEx · AI/ML Intern

Projects

Two ML systems from my PostEx internship, then school builds.

RTO risk predictor

PostEx · AI/ML Intern · June–August 2026

Two-stage LightGBM system for cash-on-delivery return-to-origin risk on real PostEx shipment data: a checkout-time model (17 features) and a post-order model (47 features), both with isotonic calibration.

Engineered five feature layers (order, customer, H3 geospatial, network/graph, temporal). Added Optuna tuning, SHAP reason codes, and champion–challenger promotion gates for retraining. Served both models through a Flask REST API with Docker and pytest. Audited explanations and fixed SHAP wording that contradicted feature values in 26% of reasons.

Stage 2 held-out test (launch thresholds cleared). Stage 1: ROC-AUC 0.90, PR-AUC 0.85.

0.92 ROC-AUC
0.91 PR-AUC
0.02 ECE
93% Precision @ 50% recall

Python, LightGBM, Optuna, SHAP, Flask, Docker, pytest

Product category classifier

PostEx · AI/ML Intern · June–August 2026

Fine-tuned a multilingual MiniLM encoder (PyTorch / Hugging Face) to predict parent category, child category, and eight operational flags — bulky, fragile, cold-chain, high-value, and more — from free-text listings against a 97-category taxonomy.

Labeled ~26,000 listings with a prompt-engineered LLM pipeline (majority voting, schema-constrained outputs), then balanced to 12,198 rows across 46 child categories. Excluded a contaminated source field that leaked labels. Selected models on child macro-F1; temperature-calibrated confidence. Served via Flask with train, retrain, batch predict, and a staging-to-production registry.

Held-out validation on 1,647 products. Balancing lifted child macro-F1 from 0.72 to 0.86.

96.2% Parent accuracy
87.7% Child accuracy
0.86 Child macro-F1
97.5% Child top-3

Python, PyTorch, Hugging Face Transformers, MiniLM, Flask, scikit-learn, Docker

Classifier demo

Synthetic listings only. Pick one to see parent, child, and flags.

Try a contour match

Synthetic demo for this page. It is not the live PostEx MiniLM model.

Parent —

Child —

Flags —

Enter a product title to run the synthetic contour match.

CyberGuard

Senior capstone · Fall 2026 – present · Team of 5 · TypeScript, React, Vite, FastAPI, Supabase

Part of a five-person team building a cybersecurity web platform with phishing email scanning and website risk tracking. I built the Cyber Assistant (2,000+ lines of TypeScript/React): a guided assistant with a 30+ topic security knowledge base, keyword and intent scoring, in-app redirects, and refusal handling for offensive-security requests.

Calendar reminder app

Course project · Fall 2025 · Grade A+ · Java, JavaFX, Firebase, MongoDB, Maven, JUnit 5

Co-developed a full-stack desktop calendar (2-person team) with Month, Week, Day, and Year views, reminder CRUD, and auth. Real-time sync via Firebase Firestore and MongoDB; JUnit 5 tests. Led sprint planning on a GitHub Projects board and designed the UI mockups in Figma.

Experience

AI/ML Intern · PostEx

June 2026 – August 2026

Built and served two ML systems on real courier data: an RTO risk predictor (two-stage LightGBM with calibration and SHAP) and a product category classifier (fine-tuned MiniLM for parent, child, and operational flags). Details and metrics are in Projects.

About

B.S. in Computer Science at Farmingdale State College (SUNY), expected December 2026. GPA 3.7 / 4.0. Dean’s List for seven semesters (Spring 2023 – Spring 2026); President’s List, Summer 2025. Based in Farmingdale, NY.

Languages: Python, Java, TypeScript, SQL, C++, COBOL
ML & data: PyTorch, Hugging Face Transformers, LightGBM, scikit-learn, MiniLM fine-tuning, LLMs & prompt engineering, SHAP, Optuna, model evaluation & calibration
Build & ship: React, Flask, REST APIs, Docker, pytest, JUnit 5, Spring Boot, Git, Figma
Data stores: SQLite, MySQL, MongoDB, Firebase Firestore