Orion: The AI-Powered Financial Conscience

Inspiration

We’ve all been there: midnight scrolling leads to a "must-have" gadget, and dopamine overrides discipline. Most financial apps are post-mortem—they tell you that you messed up yesterday. We wanted to build a pre-emptive system. Orion was born from the idea of creating a "Digital Stoic"—a system of agents that stands between your impulses and your bank account, providing a moment of high-tech reflection at the point of sale.

What it does

Orion is an AI-powered financial system that acts as your proactive conscience. By integrating directly into your shopping experience via a browser extension or screen capture, it:

Identifies: Uses computer vision to extract product details and pricing from any website or screenshot.
Analyzes: Cross-references the item against your real-time balance, monthly income, existing debts, and long-term savings goals.
Counsels: A multi-agent system debates whether the purchase is "financially sound."
Intervenes: Orion speaks to you (literally, via synthesized voice) to explain the long-term impact of the purchase on your financial health.

How we built it

Orion is built as a highly modular system of agents designed to bridge the gap between web-browsing and financial planning.

1. The Core: Multi-Agent System (LangGraph & FastAPI)

Instead of a linear script, we used LangGraph to create a directed acyclic graph (DAG) of specialized agents.

The Vision Scraper: Triggered by our bridge, it uses Gemini Pro Vision to parse product data (name, price, currency) from raw screenshots, eliminating the need for brittle CSS selectors.
The Financial Auditor: This agent queries the user's personal financial state from our FastAPI backend, analyzing the delta between current goals and the potential purchase.
The Decision Engine: A reasoning LLM that weighs the trade-offs and determines if the purchase is a "Go" or "No-Go."

2. The Bridge: Browser Extension

We developed a Browser Extension that acts as the real-time interface. It allows users to capture their current shopping cart with a single click. This bridge sends the visual data to our backend, effectively turning any e-commerce site into an Orion-compatible platform.

3. The Command Center: Glassmorphism Dashboard

The frontend is a Vanilla JS Single Page Application (SPA) designed with a modern glassmorphism aesthetic.

Personal Data Overview: Users can manage their entire financial profile, including real-time balances, income streams, and existing debts.
Goal Tracking: A dedicated module for setting and monitoring savings targets, which the agents use as "hard constraints" during analysis.
Transaction History: A clean, transparent ledger of all analyzed purchases and past spending habits stored in SQLite.

4. The Interaction Layer: Voice & Payments

Voice Synthesis: We integrated ElevenLabs to give the system a voice. The agents don't just display text; they explain their reasoning verbally to create a more impactful "conscience" effect.
Monetization: A production-ready Stripe integration handles credit purchases and Pro subscriptions, managing user quotas via secure webhooks.

Challenges we ran into

The biggest hurdle was Agentic Alignment. Initially, our agents were "too nice"—they would often approve purchases because they prioritized immediate user satisfaction. We had to implement a hierarchical LangGraph structure where a "Lead Auditor" agent could veto the "Shopping Assistant" agent.

Another challenge was latency. Running a vision model, a reasoning chain, and voice synthesis sequentially took too long. We optimized this by parallelizing the vision extraction and the financial data retrieval, cutting response time by 40%.

Accomplishments that we're proud of

Zero-UI Extraction: Successfully using Gemini Vision to identify products on any site without custom scrapers.
Functional SaaS Loop: Integrating a working Stripe credit system that makes the project feel like a market-ready product.
The "Vibe": Creating a frontend using glassmorphism that feels modern, transparent, and high-end—matching the "Orion" celestial theme.

What we learned

Orchestration > Chat: The future of AI isn't a single prompt, but a system of specialized agents working in a directed acyclic graph (DAG).
The UX of Friction: Usually, developers want to remove friction. In FinTech, intentional friction is a feature. We learned how to design AI that slows the user down just enough to make a better choice.
Real-world Reasoning: Hard-coding financial rules is brittle. Using LLMs to reason about debt-to-income ratios allows for a much more nuanced "Financial Conscience."

What's next for Turing's Grandchildren

We don't want Orion to just say "No." The next steps are:

Automated Negotiation: If Orion says "Not today," it will automatically track that item and notify you when the price drops to a level that fits your $V_p$ score.
Investment Diversion: Adding a "One-Click Save" button where, if you skip the purchase, Orion moves that money into a high-yield savings account or ETF.