Projects

A few things I’ve built outside of work — mostly experiments with LLMs, agents, and classic ML.

Bunny — a personal AI assistant

A self-hosted AI assistant I run as my daily driver. Built on Flask and the Claude Code engine, it manages my Obsidian vault, curates a news feed, plays music, tracks spending, and runs real-world errands through tool-calling sub-agents. It speaks MCP both ways — wiring in external tools as a client, and exposing itself as an OAuth remote server — and ships as an installable PWA and a Capacitor Android app.

Personal music player

A mood-tagged, Spotify-style player built over a song tracker in my Obsidian vault. Add a track by name or link and a pipeline enriches its tags with an LLM, pulls the audio with yt-dlp, and fetches time-synced lyrics from LRCLIB for a karaoke-style overlay. Flask backend, served as its own site.

Housie — a family tree

A multi-tenant family-tree web app for mapping my extended family. Supabase (Postgres + auth) on the backend, a static SPA behind Caddy, with tiered access control and a signed Android build.

Personal finance pipeline

A Dagster pipeline that turns raw bank and UPI transaction SMS into a clean ledger in Firefly III — parsing, deduping, and categorizing every transaction idempotently, with a companion job that reconciles shared expenses pulled from Splitwise.

PDF to Podcast

Inspired by Google’s NotebookLM. Extracts text from PDFs, generates a conversational script with Gemini 2.5, and converts it to speech with Edge TTS — turning any document into a two-host audio episode.

Agentic Text-to-SQL

A LangGraph-based system that translates natural language into SQL over large CPG datasets, with planning, self-correction, and validation steps so the generated queries actually run.

ChessGPT

A Streamlit app pairing Llama (via Hugging Face) with RAG over a corpus of chess books — ask it for openings, ideas, and explanations grounded in real theory.

Dockerized Time Series

An XGBoost energy-consumption predictor, fully containerized in Docker for reproducible training and inference.

YouTube Parser

A Selenium + BeautifulSoup scraper that walks a channel’s videos page and builds a structured dataset — titles, durations, upload dates, view and comment counts.