Project Thread🏆 Loophole Hackathon 2025 - Winner

GuiltSphere-AI — Agentic "Roast & Toast" Financial Auditor with Multi-Modal Vision

Overview

A chaotic financial management agent built for the Loophole Hackathon that uses emotional damage and high-energy praise to fix spending habits.

LangChain / LangGraphGoogle Gemini 1.5 Flash (Vision & Reasoning)FastAPIPostgreSQL (Vector Store)Streamlit (Mobile-Optimized Frontend)

Multi-Modal Receipt Parsing: Gemini-powered vision engine that extracts itemized data and categorizes them into "Vices" or "Virtues".
Agentic RAG Feedback: Uses long-term memory to track recurring "Financial Sins" and generates context-aware emotional roasts.
Dynamic Persona Engine: Toggleable personality modules including Disappointed Parent, Victorian Orphan, and Finance Bro.
The Wall of Shame: A social-hack feature that threatens to publish embarrassing receipts if a high Guilt Score persists for 3 days.

Implemented a "Chaos Factor" variable in the scoring algorithm to differentiate between essential spending and late-night impulse buys.

Leveraged a dual-pathway prompt architecture to ensure the AI balances humor with genuine positive reinforcement for healthy habits.

→Explored the limits of LLM "personality consistency" when handling high-variance multi-modal inputs like crumpled or oily receipts.

→Refined Agentic self-reflection loops to ensure the AI recognizes its own previous roasts to avoid repetitive dialogue.