Deployed AI Projects
Live systems you can run right now.
LLM ProtectorSelf-hosted
Automated vulnerability scanner for local Ollama models. Fires a library of categorized attack prompts — prompt injection, DAN jailbreak, roleplay bypass, system prompt extraction — and reports whether the model complied or refused each one. FastAPI backend streams results to a React/Vite frontend as NDJSON. Full WSL2 support.
Built with AI-assisted development.
AI Red-Team AgentLive
5-agent pipeline that security-tests any AI system. Describe a target — its stack, tools, and guardrails — and the pipeline maps its attack surface, generates probe attacks, audits vulnerabilities, runs adaptive mutations, and delivers a remediation report. Phase prompts are hot-swappable MD files; no redeploy needed to change agent behavior.
Data Processing AI AgentsLive
Multi-agent pipeline that automates data analysis tasks using CrewAI. React frontend on Vercel, Python agent backend on HuggingFace Spaces. Submitted to the International AI Agents Hackathon 2026.
Calendar AI AssistantSelf-hosted
Connects to Google Calendar and Gmail via OAuth 2.0. Manage your schedule in plain English — add events, query your week, turn emails into todos. 4-agent CrewAI pipeline: Intent Analyzer → Data Retriever → Processor → Verifier. Streamlit Web UI + rich CLI. Submitted to HackMars 3.0 (Neon).
Study AssistantSelf-hosted
RAG-based study tool that ingests your Markdown notes, PDFs, and images into ChromaDB and lets you query them conversationally. Runs locally via Ollama or in the cloud via Groq, OpenAI, Anthropic, or Gemini. CLI and Streamlit UI. Submitted to HackAmerica.
Data Analysis Agent ClusterLive
A crew of specialized agents that tear through data with surprising accuracy. Each agent owns one job — cleaning, prompt engineering, analysis, formatting, QA — and passes work down the chain. Feed in any document or dataset; get back structured findings. React frontend on Cloudflare Pages, Python agent backend on HuggingFace Spaces.
VillagesLive
AI-powered community learning platform where students and adult learners form small study cohorts (“Villages”). An LLM matches you to a compatible Village based on your goals and learning style. Each Village has an AI “Village Elder” that generates discussion prompts, collaborative study challenges, and structured learning paths. Real-time discussion feeds, a Topic Explorer that translates confusing material into plain language, and a full Study Hub with an AI tutor, essay coach, and study planner. Built for the USAII Hackathon 2026.
Pixel AssistantSelf-hosted
Autonomous terminal AI assistant. Complex prompts are auto-routed to ReAct-style agents that can recursively spawn sub-agents, and the assistant writes and registers its own skills at runtime. One provider-agnostic tool loop runs across a Groq → Gemini → Mistral → OpenAI → Ollama failover chain.
Pixel Code EngineSelf-hosted
Security-first AI orchestration core. The SmartEngine ranks a dynamic tool registry, plans a task into dependency-aware steps, generates new tools when none fit, and learns the winning tool-chain from every run. Secrets are scanned and redacted before any prompt reaches a model.
Currently Building
Active work-in-progress — models I'm training and iterating on right now.
tinyGPTIn Progress
A character-level GPT built from scratch in PyTorch — multi-head self-attention, stacked transformer blocks, layer norm, and a custom tokenizer, no nn.Transformer. Designed as a lightweight AI lab assistant that trains and runs on any CPU with minimal hardware, and learns from your own PDFs and text files.
MedicalAI — Light WeightResearchIn Progress
A lightweight vision-language model for medical triage. Fuses CLIP image embeddings of chest X-rays with Bio_ClinicalBERT symptom-text embeddings into a diagnosis classifier, trained on the Indiana University Chest X-ray Collection. A confidence threshold flags low-certainty cases as inconclusive so they can be routed for follow-up — my first project working with VLMs.
Classical AI & Algorithms
Game-playing agents and search algorithms built without ML libraries.
Chess AIIn Progress
Chess engine with minimax + alpha-beta pruning. Neural layer is an AlphaZero-style ResNet (policy + value heads) trained via supervised pre-training on human games, then MCTS self-play.
ML Models — What's Coming
Models I'm building to demonstrate core ML concepts from scratch. Each will have a deployed demo.
- Handwritten Digit Classifier (CNN — PyTorch)Convolutional neural net trained on 60,000 MNIST images. ~99% test accuracy after 5 epochs. Draw a digit below and the model predicts it in real time — running entirely in your browser via ONNX Runtime Web, no server needed.
PyTorch CNN MNIST ONNX onnxruntime-web Details →Draw a digit (0–9)?Loading model... - Perceptron VisualizerThe simplest neural network — one neuron, two inputs, one output. Place red and blue points on a canvas and watch the algorithm learn a decision boundary in real time. Implemented entirely in JavaScript with no ML libraries. Try the XOR pattern to see exactly why multi-layer networks were invented.
JavaScript Canvas API Perceptron Learning Rule Try it → - Transformer Architecture from Scratch (PyTorch)Multi-head self-attention, positional encoding, and a full encoder/decoder stack built in PyTorch without using
nn.Transformer. Trained on a character-level language modeling task (Shakespeare). Deployed inference endpoint on HuggingFace Spaces.
PyTorch Attention Language Modeling HuggingFace Spaces - Sentiment Analysis Fine-tune (BERT)Fine-tune
distilbert-base-uncasedon a domain-specific dataset using HuggingFace Transformers. Deploy as a REST API on HuggingFace Spaces with a React frontend for live sentiment scoring.
PyTorch HuggingFace Transformers DistilBERT Fine-tuning