Findola Platform · Knowledge graph
The enterprise knowledge graph that turns scattered work data into a living digital twin.
Cortexa is the graph layer beneath Findola. It resolves every person, document, ticket, deal, repo, and message into one connected entity graph that search and agents can reason over. "Who owns the Acme renewal across CRM, email, and docs" becomes a single graph traversal instead of a guess.
Why Cortexa exists.
Enterprise data is fragmented across dozens of SaaS tools with no shared identity model. Native search cannot answer relational questions like "which open deals depend on the stalled migration". Entity resolution at enterprise scale means millions of records, fuzzy matches, and permissions, and it breaks CPU-only pipelines.
What it does
Cortexa ingests change-data-capture streams from 200+ connected apps and runs GPU cross-app entity resolution, deduplicating and linking records that refer to the same person, account, project, or artifact across systems that never agreed on an ID. It builds a typed knowledge graph with per-edge ACLs, exposes graph queries and embeddings to the rest of the ecosystem, and keeps a temporal twin so you can ask how the org looked at any point or simulate the effect of a change.
What ships in Cortexa.
Every capability below maps to a real workload that needs accelerated compute.
GPU entity resolution
Cross-app deduplication and linkage with confidence scoring and human review.
Permission-aware graph
A typed knowledge graph with per-edge ACLs enforced at query time.
Temporal twin
Point-in-time views and what-if simulation over the graph.
Graph + embedding API
Consumed by Retrova, Agentiq, and Findola search.
Full lineage
Provenance for every resolved entity, for audit and the EU AI Act.
Architecture.
CDC streams (Kafka/Redpanda) flow into GPU parse and normalize (cuDF), then blocking and GNN linkage (cuGraph + NeMo), into a typed graph store with ACL edges, served through Triton. Findola, Retrova, and Agentiq query Cortexa. Lumind consumes its labeled links to improve the models.
NVIDIA hardware
- H100 / H200. GNN and embedding training on large graphs, infeasible on CPU.
- A100 / L40S. Production graph and entity-resolution inference.
- DGX. Periodic full-graph rebuilds across billions of edges.
NVIDIA SDKs & libraries
- RAPIDS cuGraph / cuDF. GPU graph and ETL for real-time resolution at scale.
- NeMo. Train GNN and embedding models for entity linkage on the flywheel.
- Triton. Low-latency multi-model serving of graph and embedding queries.
NVIDIA software
- NVIDIA AI Enterprise. Supported deployment in regulated and sovereign tenants.
Cortexa is GPU-essential.
Entity resolution is pairwise comparison across millions of records, quadratic without GPU blocking. GNN linkage and graph analytics are matrix and sparse operations that run 10 to 100 times faster on GPU. CPU-only cannot keep a real-time twin current or hit query latency.
The moat.
Each customer gets a proprietary resolved-entity graph that is expensive to leave, plus linkage models that compound on Lumind's flywheel. The graph gets more accurate the longer a customer runs it.
Cortexa adds a GPU-essential deep-tech core (graph ML plus entity resolution), a digital-twin and simulation narrative, and real cuGraph and NeMo usage. That is the technical depth and strategic-area fit NVIDIA reviewers reward.
Ecosystem link: Cortexa is the substrate. Retrova retrieves over it, Agentiq acts on it, Vaulta deploys it, and Lumind improves it.
See Cortexa on your own data.
Connect three apps and Findola will answer your first question in under ten minutes.