Stream Engine

The Stream Engine is the foundational intelligence layer of xFractal’s architecture. It processes heterogeneous real-time inputs from both on-chain and off-chain environments and converts them into structured token-level data points that inform agent reasoning, predictive modeling, and user-facing decision workflows. This system enables consistent, interpretable evaluation of any Solana-native asset across narrative, technical, and risk dimensions.

System Architecture

The Stream Engine comprises two independent modules: on-chain and social. Each score is computed through a dedicated data pipeline, statistical preprocessing system, and model ensemble, and then normalized for cross-comparison. The Stream Engine system is built to operate in real time, adapting dynamically to volatility conditions and token lifecycle changes.

The engine is implemented using modular MLOps architecture with support for batch and streaming inference. Data is processed through time-decayed weighting, feature scaling, statistical filtering, entity recognition, and anomaly detection routines. Custom-built models are retrained periodically using retraceable, version-controlled pipelines to ensure adaptability under evolving market conditions.

Each score is generated via a weighted composite of primary features, some derived from proprietary partner APIs, and others collected through internal ingestion tools and crawlers. The results are made available across xFractal’s natural language interface, dashboards, and trading modules.

Data Sources

xFractal integrates multiple high-fidelity data providers to ensure reliable, high-frequency inputs across different data modalities:

• Off-Chain Data (Social & Sentiment): Twitter/X (via TweetScout, Moni, ElfaAI, Masa Network, Twitter API, Twitter Scraper), Telegram (internal scrapers), Dexscreener metadata, Pumpfun data, and our proprietary, in-house-developed sentiment classification and NLP.

• On-Chain Data: Solana RPC and gRPC endpoints through Helius.

Each data stream is ingested using event-driven infrastructure and processed through custom NLP, statistical analysis, and classification models. Token tagging, contract attribution, and user behavior mapping are handled through entity linking systems trained on Solana-native datasets.

Technical Framework

The Stream Engine is supported by a scalable vectorized data infrastructure. A custom-built vector memory layer is employed for real-time context retrieval and similarity matching across historical sentiment, price, and wallet events.

Model ensembles are managed through automated pipelines, with real-time monitoring, model validation, and retraining handled via an MLOps stack. Key components are exposed through internal APIs to all downstream agents.

Security is embedded across the scoring stack via a continuous SecOps pipeline: data validation, threat detection, and access controls are enforced at ingestion and processing layers. Scoring models and outputs are stored with full traceability to ensure interpretability and compliance.

Functional Role

The Stream Engine is not a standalone analytics tool; it functions as the central intelligence source for all other xFractal components. Every agent references Stream Engine outputs when evaluating a token or determining next steps.

Typical use cases include:

• Token research and due diligence through structured, multi-dimensional scoring.

• Prompt-based intelligence queries, where the engine contextualizes responses.

• Execution filtering, enabling users to identify viable tokens based on predefined score thresholds.

• Prediction enrichment, where scoring values are used as input features for Oblivia’s model.

• Real-time dashboards for monitoring newly launched tokens, trending assets, or anomalies across sentiment and liquidity.

By centralizing raw data into structured and dynamic scoring frameworks, the Stream Engine delivers the interpretability, consistency, and scalability necessary to operate in narrative-driven, high-frequency token environments.

---