I build knowledge graphs enterprises can trust in production.

Pharmacovigilance · SHACL

Problem. Drug-safety narratives must become structured records, but a language model will produce a record that reads perfectly and is quietly wrong.

Graph decision. Make the acceptance test a SHACL validator, not another model — an agent loop extracts to RDF and repairs until the shapes graph conforms. The verifier never calls an LLM.

• RDF
• SHACL
• SPARQL
• agent loop

Security · Post-Quantum · Graph ML

Problem. Enterprises can't see which cryptographic assets are quantum-vulnerable, or who owns them, across thousands of certificates, keys, and protocols — and harvest-now-decrypt-later makes it urgent.

Graph decision. A 6,400-node crypto-posture knowledge graph with SHACL validation, PQC risk scoring against NIST/FIPS/PCI, and graph ML to rank vulnerable assets — queried in natural language via NL→Cypher, fully on-prem.

• Knowledge Graph
• Graph ML
• SHACL
• NL2Cypher
• on-prem LLM

Healthcare/Pharma · Neptune

Problem. Commercial customer data at an enterprise is scattered across 20+ systems with no shared meaning, so every question becomes a bespoke SQL request.

Graph decision. A production-shaped customer KG — Redshift → S3 → LinkML/RDF → AWS Neptune, an evolving ontology, graph analytics, GraphQL serving, and an evaluation framework to keep it honest.

• AWS Neptune
• LinkML
• RDF
• GraphQL

Privacy-first KG · Streaming

Problem. A commercial knowledge graph is only usable if sensitive identities never leak — de-identification can't be an afterthought bolted on at the end.

Graph decision. A de-identification moat (AWS Secrets Manager/KMS) wrapped around a streaming ingestion pipeline into an evolving LinkML ontology, with GraphQL, caching, graph analytics, and evaluation.

• LinkML
• KMS
• streaming
• graph analytics

E-commerce · Graph-RAG

Problem. Business teams want answers about customers in plain English, not another dashboard or a two-day SQL wait.

Graph decision. A 4,300-node / 225k-edge customer graph with PageRank, Louvain communities, and SVD embeddings, behind an NL→Cypher assistant that answers with citations.

• Neo4j
• graph embeddings
• NL2Cypher
• RAG

Data lineage · Reachability

Problem. A dropped column looks unused to dbt, then breaks a pandas script and a tablet app nobody linked to it.

Graph decision. Parse SQL and Python into one dependency graph, so "what breaks if I change this column" becomes a reachability query — statically, before the change ships.

• graph
• static analysis
• reachability

Graph-as-IR · App generation

Problem. Turning business requirements into a working app means re-modeling the same domain by hand every time — schema, API, UI — and it drifts out of sync.

Graph decision. Use a knowledge graph as the intermediate representation: extract the domain into a graph, then generate the schema, API, and React app from it, so one model drives the whole build.

• graph-as-IR
• domain extraction
• codegen
• React

On-prem · Fine-tuning

Problem. Analysts want to ask a graph questions in plain English, but the data is sensitive and can't be sent to a hosted model.

Graph decision. An ontology-grounded NL→Cypher pipeline running fully offline (Neo4j + Ollama), with a LoRA-fine-tuned Qwen2.5 for the translation, validated against a held-out query set.

• NL2Cypher
• ontology-grounded
• LoRA (MLX)
• Neo4j

Retail · Graph ML

Problem. Store × SKU demand forecasts ignore how products and stores relate — cannibalization, substitution, and shared demand shocks.

Graph decision. A graph encoder over cross-product / cross-store relationships feeds node embeddings into a global LightGBM across 4,000+ SKUs × 33 stores. ML beats naive by ~23%, with an honest read on where the graph lift stays small.

• graph encoder
• node embeddings
• LightGBM
• spatiotemporal