Agentic AI Profile — open on GitHub

A local agent that answers recruiter questions about me, and logs the lead.

A fully local Streamlit agent powered by Ollama that lets a recruiter interview my professional profile conversationally. Reasoning runs through an orchestrator that emits deterministic JSON tool calls against a structured profile store, which is what keeps it from hallucinating my background. When it detects genuine hiring intent it captures the lead to Google Sheets automatically.

• Python
• Ollama
• llama3
• Streamlit
• Tool calling
• Google Sheets API

Domain-Specific RAG Agent — open on GitHub

Production RAG with agent-controlled retrieval and an explicit refusal path.

A local-first RAG service: upload a PDF, and it chunks with metadata preserved, embeds via Sentence-Transformers, and persists a FAISS index to disk. The part I care about most is the agent layer, which decides whether to retrieve or to refuse — grounding answers in retrieved context rather than guessing when the corpus has nothing relevant. Fully offline through Ollama, exposed over FastAPI with an optional Streamlit UI.

• FastAPI
• FAISS
• Sentence-Transformers
• Ollama
• Streamlit
• Python 3.13

RAG Financial 10-K Analyzer — open on GitHub

Natural-language querying over SEC 10-K filings, entirely on local inference.

An end-to-end retrieval-augmented generation pipeline built for documents far longer than any context window: SEC 10-K filings. Handles document chunking, FAISS vector indexing and semantic retrieval, then answers with local LLM inference so nothing about the filings leaves the machine.

• FAISS
• Ollama
• LangChain
• Python
• Jupyter

LLM-Powered Sentiment API — open on GitHub

Containerized FastAPI sentiment service with CI/CD and structured exports.

A sentiment analysis service built the way a production service should be: LangChain and Ollama behind REST endpoints, structured outputs rather than free text, and automated CSV/PDF/DOCX export. Shipped with Docker and a GitHub Actions CI/CD pipeline.

• FastAPI
• LangChain
• Ollama
• Docker
• GitHub Actions

Multimodal Stress Detection — open on GitHub

CNN–BiGRU–Attention over wearable ECG, EDA and EMG signals.

A healthcare-focused deep learning model classifying physiological stress from wearable sensor data in the WESAD dataset. Built as research rather than a demo: the repo includes ablation studies, confusion-matrix analysis and research-style reporting alongside the model.

94%test accuracy

• PyTorch
• CNN
• BiGRU
• Attention
• Multimodal
• Jupyter

Circadian Rhythm Optimizer — open on GitHub

Predicts your body clock phase, then plans a 7-day realignment to solar time.

A modular deep-learning pipeline over the MMASH dataset that ingests sleep diaries, actigraphy, RR-interval and questionnaire data to estimate circadian phase. A CNN→BiGRU predicts phase encoded as (sin φ, cos φ); pvlib supplies true sunrise, solar noon and solar midnight for any coordinate, and a recommendation engine turns the phase difference into an actionable week-long light-and-sleep plan.

• PyTorch
• CNN
• BiGRU
• pvlib
• MMASH
• Python

NeuroNap — open on GitHub

Full-stack ML sleep optimization app with auth and predictive scheduling.

A complete desktop application rather than a notebook: Tkinter GUI, user authentication, SQLite-backed structured storage, and predictive scheduling models that recommend sleep windows for better quality and next-day energy.

• Python
• Tkinter
• SQLite
• scikit-learn

Streaming Churn Prediction — open on GitHub

End-to-end churn pipeline over ~210k users, explained with SHAP.

A full ML lifecycle on a multi-table streaming dataset of roughly 210,000 users: EDA, feature engineering across demographics, watch history and review sentiment, model development with hyperparameter tuning, and SHAP-based interpretability. Built to answer why users churn, not merely which ones will.

210kusers modelled

• scikit-learn
• XGBoost
• SHAP
• pandas
• Jupyter

Multi-Armed Bandits Lab — open on GitHub

Play against Greedy, ε-Greedy, UCB1 and Thompson Sampling in real time.

An interactive simulator where you compete against four bandit algorithms while the dashboard tracks cumulative reward, regret and arm-selection behaviour live. True arm probabilities stay hidden, which makes the exploration-versus-exploitation tradeoff something you feel rather than read about. Genuinely useful for teaching reinforcement learning intuition.

• Python
• Flask
• Plotly.js
• Pytest

Solar Site Mapper — open on GitHub

Geocode US addresses, map them, and pull NREL solar yield estimates.

A full-stack tool that geocodes US addresses through Nominatim, plots valid coordinates on a Leaflet map, and queries NREL solar resource and PVWatts endpoints to estimate photovoltaic output per site. Node API plus browser client.

• Node.js
• Leaflet
• NREL API
• Nominatim
• JavaScript

KitchenMind — open on GitHub

Flask app for kitchen inventory, meal suggestion and nutrition tracking.

A Flask web application that tracks what is actually in your kitchen, suggests meals from available ingredients, and handles nutrition tracking and meal-prep planning.

• Python
• Flask
• SQLite
• Jinja

NeetCode 150 Submissions — open on GitHub

Working through the NeetCode 150 in public.

My data structures and algorithms practice, committed as I go. Kept public deliberately — it is the same list I recommend to anyone on the Learn page, and it seems only fair to show my own working.

• Python
• Algorithms
• Data Structures