Heabsy
Heabsy provides ARGUS, an enterprise knowledge infrastructure combining knowledge graphs with autonomous agents to deliver auditable, multi-hop reasoning for complex document intelligence. The system runs entirely on customer-controlled hardware within the EEA, achieving 56.7% exact match on the MuSiQue benchmark versus roughly 20% for naive RAG. It targets DACH-based enterprises in regulated industries such as legal, finance, and healthcare.
- Artificial Intelligence
- AI Agents
- Enterprise Software
- Data & Analytics
- Software Only
Funding
Founders
Founder details are not available yet.
Product
Problem
Standard retrieval-augmented generation (RAG) systems fail at complex, multi-step questions in enterprise settings—they hallucinate facts, miss connections across documents, and provide answers without explanations. This limits AI adoption in regulated industries where accuracy and traceability are critical.
Solution
ARGUS is an enterprise knowledge infrastructure that combines a knowledge graph with an autonomous agent to simulate human expert problem-solving. Instead of relying on simple vector similarity, ARGUS indexes documents into 260,000+ knowledge graph triplets, enabling multi-hop reasoning that connects information across sources. The graph-based filtering approach allows a smaller open-weight model to match the output of much larger models, making on-premise inference affordable. The system provides auditable, explainable answers, and deployments run entirely on customer-controlled hardware inside the EEA, ensuring GDPR and EU AI Act compliance.
Target Audience
Primary customers are DACH-based enterprises in regulated industries (legal, finance, manufacturing, healthcare) that need secure, auditable AI for document intelligence and complex question-answering.
Features
- Knowledge graph construction from documents with entity resolution improving graph connectivity from 1% to 46%
- Multi-hop reasoning over knowledge graph triplets for questions that require connecting multiple documents
- Autonomous agent architecture that investigates and iterates on queries rather than performing single-pass retrieval
- On-premise deployment on customer infrastructure within the EEA with SOC 2 roadmap
- Benchmarked performance: 56.7% exact match on MuSiQue, 84% exact match on 2WikiMultiHopQA, 65% exact match on HotpotQA, tested on 500+ samples per benchmark
- Production-scale indexing of 19,000+ documents with 260,000+ knowledge graph triplets