Feroz Khan - Portfolio

Fine-Tuning CNN Architectures Using Transfer Learning

Baseline CNN → VGG16/ResNET fine-tuning → +4.5% accuracy (91% → 95.5%).

Fine-tuned pre-trained VGG16 & ResNet on Fashion MNIST using a custom PyTorch Dataset + ETL pipeline.
Frozen backbone, replaced classifier head, tuned with K-Fold + Optuna for generalization.

PyTorch Transfer Learning Optuna Computer Vision

Code · Documentation

Sentiment Analysis for Healthcare Domain with NLP & MLOps

Mental health text classification → 80% accuracy · 0.84 ROC-AUC · 12% F1 disparity reduction.

Built end-to-end NLP pipeline on ~15K patient text records using Naive Bayes and Logistic Regression.
Integrated MLflow experiment tracking and evaluation pipeline.

NLP MLflow Scikit-learn SQL Responsible AI

Code · Documentation

Agentic RAG with Hallucination Filtering & Self-Reflection

Hallucination-prone RAG → Agentic retrieval + grading → +15% GPT-Judge accuracy · 20% redundancy reduction.

Built modular LangGraph-based RAG with MMR + multi-query retrieval. Dockerized and deployed on AWS with CI/CD and FastAPI endpoint.

LangGraph FAISS Pydantic CI/CD Docker

Code · Documentation

Fine-Tuning LLM with Quantization & LoRA

Generic support replies → Gemma-7B + LoRA (8-bit) → +14% relevance · 60% lower memory.

Fine-tuned Gemma-7B on 945K+ customer support tweets using PEFT LoRA adapters and 8-bit quantization. Built HF Trainer pipeline to train, checkpoint, and load adapters for GPU inference.

PEFT LoRA Quantization Transformers bitsandbytes LLMs

Code · Documentation

End-to-End Sales BI Pipeline & Dashboard

Raw OLTP data → Star schema + automated ETL → 20% reduction in ad-hoc analysis.

Designed scalable fact-dimension model and processed 150K transactional records using SQL, Power Query, and Airflow. Built Power BI dashboards with KPIs and decomposition trees for drill-down revenue analysis.

MySQL Airflow Power BI Star Schema ETL

Code · Documentation

IMT 526: Building & Applying LLMs

LLM fundamentals → 15+ lab assignments (N-grams → Transformers → RAG) → reusable PyTorch implementations.

A collection of Jupyter notebooks and assignments from the LLM course I took from Prof. Chirag Shah at the University of Washington. Topics included language modeling, tokenization, embeddings, RNN/LSTM, attention, fine-tuning, and evaluation.

Repository

End-to-End Regression with AWS CI/CD

Overfit-prone regression → L1/L2 + feature selection + CI/CD deploy → MAE 3.8 · RMSE 5.1.

Built modular training + inference pipeline (ingestion, transform, train) with regularization and tuning to improve generalization. Dockerized and deployed on AWS using ECR + EC2 with GitHub Actions CI/CD.

Code · Documentation

Writing & Notes

Technical deep dives on ML systems, Deployment, LLM safety, and production AI workflows.

Topics include L1/L2 and Batch Normalization, MLOps workflows, AI Agents, model evaluation and metrics, etc.

Visit Medium