Home page
Home page - 2
Agent Drawer
Telemetry log
Notification Drawer
Market Assistant
Portfolio Page
Watchlist & Alerts
Stock Workspace
Stock Workspace -2
Stock Workspace -3
Agent Status details page
Agent Status details page -2
Agent Status details page -3

Inspiration

Retail financial dashboards are fundamentally broken. They either overwhelm users with static, lagging charts that require manual calculation, or offer simple generative AI chats that fabricate statistics, suffer from look-ahead bias, and hallucinate recommendations.

We were inspired to build a true institutional-grade Collaborative Partner for Indian retail equity-market participants (NSE / BSE). Our goal was to create an autonomous, self-correcting research desk that runs 15 specialized analyst agents in parallel. Aureon provides a collaborative trading assistant that does the heavy lifting: observing the market, running quantitative pivots, conducting point-in-time analogue audits, challenging its own bullish theories, and presenting auditable evidence directly to the user for one-click approval.

What it does

Aureon Intelligence operates in two dynamic modes to assist traders continuously:

Live Trade & Ingestion Mode: Observes market activity. Users can ingest their portfolio in natural language (e.g., "I own 100 shares of Reliance at ₹2,450"), which is parsed by our conversational assistant into structured database proposals. The quantitative layer monitors support and resistance boundaries, volume z-scores, and sector relative strength.

Guardian Creation: Once the user ingests its portfolio, Guardians are created for each stock which have access to all the 15 agents present in the application. All these guardians monitor there respective stocks parallelly.

Autonomous Multi-Agent Investigation: When a technical breakout or signal is detected, the Chief Intelligence Orchestrator (CIO) dispatches a task to the background worker pool. Specialist agents (Technical, Regime, News, Risk) analyze the security. Crucially, a Devil's Advocate agent is dispatched to find counter-arguments and challenge the bullish thesis, ensuring the final report is balanced and free of bias.

Point-in-Time Analogue Memory: Aureon matches the current market state vector against a database of 10,000+ historical records. It extracts matching states strictly at historical point-in-time timestamps T with zero look-ahead bias to calculate objective win rates, median returns, and adverse excursions mathematically.

How we built it

We engineered Aureon using a fully modern, production-grade Google Cloud stack:

Agent Orchestration: Built entirely using the Google Agent Development Kit (ADK) and the new Google GenAI SDK (google-genai). The core reasoning loop runs on Gemini 3.5 Flash.

Backend Architecture: Formed as a stateless Python FastAPI application containerized via Docker and deployed on Google Cloud Run. The real-time telemetry feed operates over persistent WebSockets (/api/v1/stream) to broadcast live ticks and background agent activity logs.

Data & Storage: Powered by a Google Cloud SQL PostgreSQL instance, utilizing a Short-Lived SQLAlchemy Session registry to isolate user portfolios, signals, active theses, and historical datasets.

Presentation Layer: Built on Angular 20 (Standalone SPA) and served via Firebase Hosting for low-latency distribution.

Challenges we ran into

Look-Ahead Bias Elimination: Calculating historical analogues is prone to leakage (using future data at past timestamps). We solved this by developing a strict point-in-time feature extraction model. Features and outcomes are stored in separate, isolated tables linked by exact historic timestamps, ensuring the similarity matcher only queries data available at the moment of the signal.

Cloud SQL Secret Integration: Packaging database credentials securely without hardcoding was critical. We resolved this by integrating standard Google Secret Manager environment overrides in our FastAPI database connection factory.

Accomplishments that we're proud of

Build: We were able to build the platform that we thought as an idea into actual product with the help of the resource we got from the google platforms.

What we learned

Memory vs. Persistence: In agentic workflows, persistence is not the same as memory. Dynamic memory requires vector-matching point-in-time states.

Collaborative Self-Correction: Agents are much more effective when organized with conflicting objectives. Spawning a "Devil's Advocate" agent to challenge the "Technical Analyst" agent dramatically reduces confirmation bias in stock research.

What's next for AUREON INTELLIGENCE

Global Idea: Since we are am from India, we currently only have the knowledge of how Indian Equity markets work. Once Aureon gets stable with the Indian market, we will be adding the global stock equity market data also into it, making the platform available to all the investors across the globe.