Our Team

Our team combines over a decade of collaborative experience in machine learning, graph science, and full-stack development. Prior contributions include unsupervised learning tooling in MeTTa, GraphRAG integrations with Neo4j, and open-source agent frameworks.

Company Name (if applicable)

Project details

Introduction

Knowledge graphs encode concepts as nodes and their interrelations as edges, forming a structured substrate for reasoning in LLM-powered applications [1]. However, assembling high-quality KGs from unstructured text remains challenging: errors in extraction, noisy or redundant connections, and LLM hallucinations hinder trustworthiness [2]. Our goal is to bridge these gaps by creating a unified framework that (a) constructs KGs automatically from scientific publications, (b) refines and analyzes graph structures, and (c) grounds LLM outputs through graph-based retrieval and reasoning.

Project Details

• Technical Approach
  The system architecture comprises four components:

  1. KGTool: Core library offering graph construction, transformation, and query routines.

  2. KG Studio Backend: REST APIs exposing KGTool functionalities, orchestrating workflows for ingestion, analysis, and knowledge-grounded generation via the GraphRAG paradigm [3].

  3. KG Studio Front end: Web interface enabling non-technical users to upload documents, explore and manipulate KGs, and visualize reasoning paths.

  4. Graph Storage: Flexible backend (e.g., Hyperon) for persisting and indexing graph data.

• LLMs will drive:

  1. Entity and relation extraction from scientific texts.

  2. GraphRAG: subgraph retrieval and augmentation of LLM responses with explicit graph citations.

• Methodology

  1. Extraction

     • Identify concept nodes and typed edges via prompt-based LLM calls.

     • Allow optional specification of entity and relationship types for domain adaptation.

  2. Graph Refinement

     • Denoise: compute centrality and degree metrics to prune low-importance nodes.

     • Distill: apply spanning-tree algorithms to preserve core connectivity.

     • Cluster: detect communities and summarize clusters into higher-level concepts via LLMs.

  3. Conflict Resolution

     • Merge duplicate entities/relations and resolve logical inconsistencies (e.g., contradictory “IS” assertions) through rule-based and embedding-based alignment routines.

  4. Fact-Grounded Generation [3, 4, 5]

     • Embed graph contents to support retrieval-augmented queries.

     • Implement GraphRAG to match LLM-suggested subgraphs with stored KG facts, ensuring citations of source publications.

  5. Visualization & Interaction

     • Front-end dashboards for graph exploration, node/edge attribute inspection (including confidence scores), and stepwise control over workflows.

     • Visual explanations of reasoning chains, community structures, and grounding evidence.

References

1. Hogan A. et al. “Knowledge graphs.” ACM Comput. Surveys 54, no. 4 (2021): 1–37.
2. DeLong L.N., Fernández R., Fleuriot J.D. “Neurosymbolic AI for reasoning over knowledge graphs: A survey.” IEEE Trans. Neural Nets Learn. Syst. (2024).
3. Edge D. et al. “From local to global: A graph RAG approach to query-focused summarization.” arXiv:2404.16130 (2024).
4. Luo L. et al. “Reasoning on graphs: Faithful and interpretable large language model reasoning.” arXiv:2310.01061 (2023).
5. Sun J. et al. “Think-on-graph: Deep and responsible reasoning of large language model on knowledge graph.” arXiv:2307.07697 (2023).

Open Source Licensing

Background & Experience

• Open‑source projects integrating GenAI, knowledge graphs, and agents. 
• Enterprise solutions for aviation, FMCG, finance, consultancy, and law.
• Entropy-weighted pooling for GNNs [1] and topology-aware sampling for graph classification [3].
• Contrastive representation learning for dynamic link prediction [2].
• Boundary-informed inverse PDE solvers and optimal control on graphs [4][5].
• Self-supervised GNNs for community detection [6].

1. “Maximum entropy weighted independent set pooling for graph neural networks.” (2021).
2. “Contrastive representation learning for dynamic link prediction in temporal networks.” (2024).
3. “Topology-aware graph signal sampling for pooling in graph neural networks.” (2021).
4. “Inverse boundary value and optimal control problems on graphs: A neural and numerical synthesis.” (2022).
5. “Boundary informed inverse PDE problems on discrete Riemann surfaces.” (2022).
6. “Graph representation learning in a contrastive framework for community detection.” 2021.

Links and references

• https://github.com/SNetGrant/metta-ul
• https://github.com/vhybzUX/confidant
• https://github.com/vhybzUX/lattice