An AI-led medical review assistant for pharmacovigilance workflows. It reads an unstructured clinical safety narrative and returns a structured assessment covering seriousness, MedDRA PT/SOC coding with 8-digit codes, labelling status, and WHO-UMC causality.

The repo builds a complete training pipeline from openFDA FAERS reports, MedDRA lookups, FDA drug label text, Qwen3-32B teacher generation, and Qwen3-14B LoRA fine-tuning. It also includes inference, evaluation, and a Gradio demo interface.

Evaluation showed 94% seriousness accuracy, 94% MedDRA PT accuracy, 98% SOC accuracy, 92% labelling status accuracy, and 80% all-fields-correct performance, a +63.4 percentage point lift over the base model.

Source code is available on GitHub.

friend.ly is a full-stack social media platform built with Django, Django REST Framework, SQLite, and a JavaScript frontend. Users can create image posts, follow profiles, like and comment on posts, search for people, and browse paginated feeds.

Its recommendation engine builds a directed social graph with NetworkX, finds second-degree connections, and combines mutual-follower influence, Jaccard similarity across liked posts, and profile-search history to rank people a user may know.

The backend includes custom user, follow, post, like, comment, and search-history models plus API routes for the interactive social features.

Source code is available on GitHub.

SecIoT is a full-stack interface for comprehensive security and forensic analysis of smart-home and healthcare IoT devices and wearables. It separates workflows for networked devices from analysis of the mobile apps used to control them.

The React client provides dedicated flows for Nmap port discovery, OpenVAS vulnerability assessment, Metasploit exploitation, SPIKE fuzzing, MobSF mobile security analysis, and APKTool reverse engineering. Each workflow captures the inputs needed to dispatch its assessment job.

A Node.js and Express service backed by MongoDB handles account registration and sign-in, bcrypt password hashing, session state, and JWT-based authentication.

Source code is available on GitHub.

Nim-AI is a Python implementation of the mathematical strategy game Nim, paired with an agent that learns entirely through reinforcement learning and self-play.

The agent stores Q-values for state-action pairs and updates them from win, loss, and intermediate rewards. An epsilon-greedy policy balances exploration with the best-known move during training, while gameplay switches to pure exploitation when facing a human.

The included runner trains the agent across 10,000 simulated games before launching an interactive command-line match.

Source code and a gameplay demo are available on GitHub.