Over the past year, the Propheus Research team has been developing an AI system designed to understand and interpret the real world.

How can enterprises make sense of the world around their business—not just what's in their systems, but what's outside their walls?

ALCHEMY: Augmented Learning and Contextual Harnessing for Enterprise Models and Yield

1. Augmented Learning – Leveraging LLMs with structured + unstructured external knowledge sources

2. Contextual Harnessing – Using real-time and contextual signals (e.g., location, time, events, sentiment, etc.)

3. Enterprise Models – Tailored AI/ML models optimized for real-world business problems

4. Yield – Optimizing ROI, accuracy, and decision-making outcomes for enterprises

   
   

Alchemy was designed with the ability to understand the underlying semantic context from data—rather than relying on pattern detection or factual answering. It can answer the what, why, and how behind enterprise first-party data and real-world signals.

Architecture 

Propheus LLM

The core AI engine that understands user prompts and coordinates actions across agents and data sources.

Agents

They parse the user’s query to understand intent and extract key entities, then classify it into types like factual, pattern-based, reasoning, or predictive—routing it to the appropriate expert logic.

Experts

Task-specific models and logic embedded in each Agent to deliver accurate, domain-relevant outputs (e.g., Lease Abstraction Agent, Product Allocation Agent).

Data

Propheus fuses first-party data with its Digital Atlas—a multimodal layer of real-world signals like mobility, demographics, POIs, and satellite data—feeding it into the Semantic Data Kernel, which transforms raw tables into business concepts for downstream reasoning.

Journey of a Query Through Alchemy

'Concepts', a critical component of the Semantic Data Kernel, themselves are defined as an abstract mental representation of a domain. They distill and organize fundamental characteristics, capturing essential attributes and relationships while filtering out unnecessary details. This process makes information more accessible to Large Language Models (LLMs), enabling clearer understanding and analysis.

To assess the accuracy of a query’s journey from input to final result, we tested Alchemy against RAG-based models on a set of retail-specific domain questions. One key observation was that Alchemy outperforms RAG by ~40% in composite accuracy score, based on tests across 22 models.

Here are some other observations.

How does query handling differ in Alchemy versus Standard RAG Pipelines?

We’ve benchmarked Alchemy across a range of state-of-the-art (SOTA) models - which we’ll explore in detail in our upcoming blog.

While model size and capability remain important factors, the architectural design of the AI framework plays a critical role in overall performance and reliability. The findings highlight the potential of agentic frameworks like Alchemy to handle complex data interaction tasks more effectively.

Contact us to learn more!