The Forge

Explainable ML

Explainable Eye Disease Detection

This project detects eye diseases from retinal images, and — just as important — shows why it made that call. Every prediction comes with a heatmap of what the model was actually looking at, not just a label. It covers four diagnostic categories, and the whole point was to treat explainability as something the model has to earn, not a nice extra tacked on afterward.

• PyTorch 2.0+
• timm (pretrained architectures)
• Grad-CAM
• Gradio
• Hugging Face Spaces

Trained on ~4,200 retinal images across four diagnostic categories (Kaggle's Eye Diseases Classification dataset).

NEEDS: real prediction + Grad-CAM heatmap image pairs for a few example cases, to extend this deep-dive with an actual raw-prediction vs. Grad-CAM overlay toggle.

LLM systems

AutoPrompt v2.0

AutoPrompt automatically improves prompts for an LLM instead of me tweaking them by hand. It runs on Groq's API using Llama-3.1-8B, and the idea is simple: instead of guessing which phrasing works best, let a separate process generate variations and actually test them. It grew out of wanting prompt engineering to be less trial-and-error and more of a measured process.

A separate model generates and tests candidate prompts against the task. Baseline and optimized versions run at the same time using asyncio, with rate limits managed through semaphores so nothing gets throttled or dropped. Both versions get scored against held-out test data, so the better prompt is picked on actual results, not a guess — and running everything concurrently is also what makes the whole pipeline noticeably faster than testing prompts one at a time.

This work also became a co-authored, accepted paper, which was a good lesson in the difference between a tool that works and a result that holds up to review.

• Python · asyncio
• Groq API
• Llama-3.1-8B
• Semaphore-managed concurrency

Full-stack ML

Second-Hand Price Intelligence

This tool estimates a fair resale price for second-hand items. A scraper pulls real listings, an XGBoost model predicts the price, and the whole thing is wrapped in a proper backend and dashboard instead of living in a notebook. The goal was to build something a person could actually open and use to check what their stuff might sell for, not just a proof-of-concept script.

FastAPI serves the backend, a Streamlit dashboard is the frontend, and everything runs containerized with Docker — a complete path from raw listing data to a usable price estimate. The scraper handles data collection on its own, feeding real listings into the model instead of relying on a static dataset, so the predictions stay grounded in what's actually being sold.

• XGBoost
• FastAPI
• Streamlit
• Web scraping
• Docker

NEEDS: real CI-score / SHAP output examples, to extend this piece with a live prediction-breakdown interaction like the brief originally envisioned.

Applied ML

Diabetes Prediction

This project compares a batch of different machine learning algorithms to see which one actually predicts diabetes best from patient health data — glucose levels, BMI, blood pressure, age, and a few other clinical measurements. Instead of picking one model and going with it, the idea was to run many side by side under the same conditions and let the results decide. It was a shared project with a teammate, and each of us tried a different way of preparing the data before classification.

My part used PCA for dimensionality reduction paired with a kernel SVM (RBF kernel), run against the standard eight-feature diabetes dataset — Pregnancies, Glucose, Blood Pressure, Skin Thickness, Insulin, BMI, Diabetes Pedigree Function, and Age — across a few different train-test splits. The rest of the project tested a much wider set of algorithms, from boosting methods to neural nets, on the same data, which made it a good way to see how much the preprocessing choice alone changes the outcome.

• Python
• scikit-learn
• pandas / numpy
• PCA
• Kernel SVM (RBF)