saaz.dev@portfolio:~$ ls projects/featured

GitHub work and flagship builds.

This page pulls from both the latest resume and the most relevant repositories on my GitHub. The focus is on systems that were built with clear technical intent, not just one-off demos.

fraud_detection/
├── api
├── mlflow
├── dvc
└── monitor

Fraud Detection MLOps Pipeline

Production-oriented fraud detection system built on the IEEE-CIS dataset with XGBoost, experiment tracking, versioned data workflows, monitoring, CI, and API serving.

XGBoost MLflow DVC Evidently AI

Scale: 590k transactions from IEEE-CIS.
Result: ROC-AUC 0.93 with 47 passing pytest tests.

GitHub repo

governance/
audit.py
rules.yaml
api/
reports/

Model Card Compliance Auditor

An AI governance tool for evaluating whether model cards are complete, compliant, and usable by downstream teams. Built as a practical quality layer around documentation and review.

Compliance FastAPI AI Governance

Goal: improve model documentation quality before models move deeper into delivery workflows.
Angle: treats governance as a product and tooling problem, not a PDF problem.

GitHub repo

retriever = FAISS()
embed = MiniLM()
query -> top_k
score = 5/5

simple-RAG

A focused retrieval experiment built around sentence-transformers and FAISS, with attention on retrieval quality and lightweight evaluation instead of hype-only demos.

Sentence Transformers FAISS Evaluation

Signal: achieved 5/5 retrieval accuracy in the documented evaluation flow.
Use: forms the kind of retrieval foundation I build on for larger assistant products.

GitHub repo

EffNetB3 + attention
Grad-CAM
5-class grading
< 20MB edge target

SA-EffNetB3 for Diabetic Retinopathy

Attention-guided EfficientNetB3 pipeline for 5-class diabetic retinopathy grading on Messidor-2, built to balance medical-imaging performance with deployment practicality.

Computer Vision PyTorch Grad-CAM

Result: 95.77% validation accuracy with fewer than 5M parameters.
Engineering detail: augmentation, oversampling, and explainability built into the workflow.

GitHub repo

FastAPI
Gemini AI
scrapers/
prompt ops

Memenem Backend

An AI-powered backend for meme generation using FastAPI, Gemini, and multi-source scraping. A playful product, but built with real backend structure and API thinking.

FastAPI Generative AI Backend

Focus: orchestrating model calls and content sources behind a clean service layer.
Why it matters: not every AI build has to be serious to still prove strong engineering habits.

GitHub repo

MemoryPal
summary
flashcards
Q&A
SQLite SRS

MemoryPal AI Study Assistant

A resume-highlighted RAG-style study assistant that turns notes into summaries, flashcards, and Q&A while tracking spaced repetition with SQLite. Designed for repeated use, not one-time prompting.

RAG System FAISS SQLite Streamlit

Core idea: use retrieval plus spaced repetition to make study sessions more useful over time.
Source: featured in the latest resume as a flagship AI build.

Public link available on request