Introduction
ORSTAC has deployed an autonomous AI Cognitive Agent that transforms how dev-traders interact with the world’s largest open-source trading bot repository. Built on a modern RAG (Retrieval-Augmented Generation) stack powered by n8n, PostgreSQL with pgvector, Google Gemini, and DeepSeek, this agent has ingested all 3,360 trading bots in XML Blockly format, the complete Deriv WebSocket API specifications, and quantitative strategies extracted from the ORSTAC blog into a semantic vector database with HNSW indexing. The result is an intelligent assistant that can analyze bot architectures, recommend optimal configurations based on market conditions, and continuously learn from live trade execution telemetry.
This release represents a paradigm shift from static bot repositories to living, adaptive trading intelligence systems. The agent doesn’t just store bots — it understands their structural DNA, maps their strategy archetypes, and correlates their performance characteristics with real-time market regimes. Join the community discussion on Telegram and explore the full agent stack on GitHub. For hands-on implementation of the strategies the agent recommends, start with a demo account on Deriv.
Trading involves risks, and you may lose your capital. Always use a demo account to test strategies.
Architecture: n8n + pgvector + Dual LLM Stack
The ORSTAC Cognitive Agent runs on a containerized Docker Compose stack combining n8n as the AI-native workflow orchestrator, PostgreSQL with pgvector for vector similarity search, and a dual-LLM configuration using Gemini as the primary reasoning engine with DeepSeek as an automatic fallback. This architecture was designed for resilience — if Gemini hits rate limits or API downtime, DeepSeek seamlessly takes over without interrupting the agent’s cognitive pipeline.
The pgvector extension enables HNSW (Hierarchical Navigable Small World) indexing on 1536-dimensional embeddings, providing sub-millisecond approximate nearest neighbor searches across the entire knowledge base. This means when a trader asks “which bot works best for Volatility 100 Index with RSI confirmation,” the agent performs a semantic search — not keyword matching — across all 3,360 bot descriptions, their XML block structures, and related blog strategy articles simultaneously.
The n8n orchestration layer manages three core workflows: `01ingestresources` for initial RAG ingestion and vectorization, `02cognitivebrain` for interactive chat with full retrieval-augmented context, and `03learningloop` for continuous feedback processing. Each workflow is modular, version-controlled, and exportable as JSON blueprints. Explore the complete architecture and deployment instructions on GitHub. For traders ready to deploy the strategies surfaced by the agent, Deriv provides the DBot platform for loading XML bots directly.
Dr. Ernest Chan, in his foundational work on quantitative trading, emphasizes the critical role of systematic knowledge management in algorithmic strategy development:
“The edge in quantitative trading increasingly comes not from a single model but from the systematic organization and retrieval of market knowledge — the ability to rapidly connect historical patterns with current conditions and execute accordingly.” GitHub
Forensic Bot Mapping: 3,360 Bots Analyzed and Categorized
The agent’s knowledge base includes a complete forensic audit of all 3,360 XML trading robots, categorized by market distribution, contract type, technical indicators, and money management models. This structured mapping enables the agent to make informed recommendations based on empirical data rather than generic advice.
The forensic analysis reveals that 59.1% of bots (1,985) target the Volatility 100 Index — the most active synthetic market preferred for rapid Digits and Martingale strategies. Contract type distribution shows Rise/Fall dominating at 33.9% (1,139 bots), followed by Over/Under at 21.1% (710 bots) and Matches/Differs at 13.6% (457 bots). For technical indicators, RSI leads with 459 bots, followed by SMA (417), MACD (227), Bollinger Bands (179), and EMA (143).
Perhaps most critically, the money management analysis reveals that 30.6% of bots (1,030) use Martingale/multiplicative recovery — a strategy that requires careful risk management due to its exponential drawdown potential. Only 3.3% (110 bots) use Soros/compounding, while 66.1% use flat stake or custom progressive lists. The agent uses this distribution data to warn traders about risk profiles and suggest position sizing adjustments based on the Kelly Criterion adapted for non-stationary markets.
Marcos Lopez de Prado’s research on financial machine learning directly supports this forensic approach to strategy categorization:
“Strategy taxonomy is not merely an organizational exercise — it is a prerequisite for portfolio construction. Understanding the structural similarities and differences between strategies prevents over-concentration in correlated approaches and enables true diversification of alpha sources.” GitHub
XML Structural Analysis: How the Agent Reads Bot DNA
The cognitive agent parses Deriv DBot XML at the block level, mapping the four-stage execution pipeline: trade definition, beforepurchase analysis, purchase execution, and afterpurchase capital management. This structural understanding allows the agent to explain exactly why a bot behaves the way it does, identify potential weaknesses, and suggest modifications.
The XML parsing engine detects critical elements: `SYMBOLLIST` fields for market/asset identification (`R100` for Volatility 100), `TRADETYPECATLIST` and `TRADETYPELIST` for contract configuration, indicator blocks (“, “, “) for technical analysis logic, and mathematical arithmetic blocks in the `afterpurchase` section for money management detection. When the agent identifies a Martingale multiplier — typically a `matharithmetic` block with `MULTIPLY` operation following a loss check — it automatically flags the risk profile and calculates the probability of ruin using the formula $P_{loss}(k) = (1-p)^k$ for $k$ consecutive losses.
This deep structural parsing transforms the agent from a simple search tool into a genuine bot architect. A trader can describe their desired strategy in natural language — “I want a mean-reversion bot for Volatility 75 with Bollinger Bands entry and flat stake management” — and the agent will search its vector database for the closest structural matches, explain the differences between candidates, and suggest parameter adjustments based on current market volatility regimes.
Continuous Learning Loop: From Trade Telemetry to Episodic Memory
The `03learningloop.json` workflow exposes a webhook endpoint that receives live trade execution telemetry — payout results, loss sequences, martingale step counts, and market volatility measurements — and processes them into long-term episodic memory for future recommendations. This creates a feedback loop where the agent’s advice improves with every trade executed by the community.
When the webhook receives telemetry data, the agent performs three operations: (1) logs entry variables and results into a `learning_logs` table for historical analysis, (2) runs anomaly detection on loss sequences — if consecutive losses exceed a configurable threshold, it triggers an LLM analysis of what went wrong under current market conditions, and (3) saves the generated insight as a vector-embedded episodic memory entry. This means the next time a trader asks about a strategy that previously caused anomalous losses under similar market conditions, the agent will proactively surface that historical insight.
This architecture implements what Dr. Andrew Ng describes as the “data flywheel” pattern for AI systems — each interaction and trade result makes the system incrementally smarter. For the ORSTAC community, this means the collective trading experience of thousands of dev-traders is distilled into actionable intelligence accessible to everyone.
“The most impactful AI systems are not those with the largest models, but those with the tightest feedback loops between predictions and outcomes. A system that learns from its own deployment data compounds its advantage over time in ways that static models cannot match.” GitHub
Comparison Table: ORSTAC Agent Stack Components
| Component | Role | Technology |
|---|---|---|
| Workflow Engine | Orchestrates RAG pipelines, chat, and learning loops | n8n (AI-Native Runner) |
| Vector Database | Stores and retrieves semantic embeddings with HNSW indexing | PostgreSQL + pgvector v0.5.1 |
| Primary LLM | Logical reasoning, prompt inference, response generation | Google Gemini API |
| Fallback LLM | Redundancy for rate limits or primary endpoint downtime | DeepSeek API |
| Knowledge Sources | 3,360 XML bots, Deriv API specs, blog strategies, bot mapping | Markdown + JSON documents |
| Learning Layer | Webhook-driven trade telemetry processing and episodic memory | n8n workflow + pgvector |
Frequently Asked Questions
What is the ORSTAC AI Cognitive Agent?
The ORSTAC AI Cognitive Agent is an autonomous RAG-powered assistant that has ingested the complete ORSTAC repository — 3,360 trading bots, API documentation, and quantitative blog strategies — into a semantic vector database, enabling intelligent search, bot analysis, and strategy recommendations through natural language interaction.
How does the agent differ from a simple search over bot files?
The agent performs semantic similarity search using vector embeddings rather than keyword matching. This means asking “show me conservative strategies for trending markets” will return relevant bots even if they never use the words “conservative” or “trending” — the agent understands the conceptual meaning behind bot architectures and matches intent to structure.
Can the agent learn from my trading results?
Yes, the continuous learning loop accepts trade execution telemetry via a webhook endpoint. When you send payout results, loss sequences, and market conditions, the agent processes anomalies, generates insights using LLM analysis, and stores them as episodic memory that improves future recommendations for the entire community.
What infrastructure do I need to run the agent?
You need Docker and Docker Compose installed on any machine with at least 2GB RAM. The agent stack runs two containers — PostgreSQL with pgvector and n8n — and requires API keys for Gemini (primary) and optionally DeepSeek (fallback). Full deployment instructions are in the `orstac-agent/README.md`.
Is the agent open-source?
Yes, the entire agent stack is fully open-source and available in the ORSTAC repository under the `orstac-agent/` directory. All n8n workflows are exported as JSON blueprints, the knowledge base documents are in Markdown, and the Docker Compose configuration is ready for one-command deployment.
Conclusion
The ORSTAC AI Cognitive Agent marks a fundamental evolution in how open-source trading communities organize, access, and leverage collective knowledge. By combining RAG architecture with structural XML analysis and continuous learning from trade telemetry, the agent transforms a static repository of 3,360 bots into a living intelligence system that grows smarter with every interaction.
The dev-trader community now has access to an AI assistant that understands bot DNA at the block level, correlates strategy archetypes with market regime data, and proactively surfaces insights from the community’s collective trading experience. This is not incremental improvement — this is the infrastructure for the next generation of algorithmic trading intelligence.
Explore the full agent stack, deploy it on your own infrastructure, and start contributing to the collective intelligence at Orstac. Join the discussion at GitHub. Start implementing the agent’s recommendations with a demo account on Deriv.
Trading involves risks, and you may lose your capital. Always use a demo account to test strategies.