PII-safe synthetic corpora
Training data generated from your own graph structure, so it's useful without exposing a single real record.
Synthos · The simulation layer
A digital twin of your org that manufactures safe training data.
Synthos is Entilla's simulation and synthetic-data engine. It lets institutional memory be simulated and hardened, not just stored, so new customers start warm and every model gets tougher evals.
The problem
Cold graphs, locked data, hidden risk.
New customers start with an empty graph and no labeled data. Privacy rules forbid training on real PII. And most orgs can't see their knowledge risk until a key person walks out the door. There's no safe way to generate training data or to ask "what if this team left?"
What Synthos does
Simulate the org, then generate safe data from it.
Synthos builds a simulation of org structure, workflows, and decision flows from the Graphos graph, then generates synthetic corpora and edge cases with no real PII exposed. That bootstraps new customers, hardens Verix evals, and trains Mnemo and Nexora models. It also runs scenario simulations for attrition, reorg, and knowledge decay to surface risk early.
Core features
What's inside.
Organizational twin
Structure, workflows, and decisions, modeled.
Scenario simulation
Attrition, reorg, and knowledge-decay runs.
Edge-case generation
Hardens Verix evals against rare failures.
Cold-start bootstrap
Seed datasets for brand-new deployments.
Knowledge-risk dashboards
See where critical context is concentrated.
Why it needs GPUs
Generation and simulation at org scale.
NeMo
Synthetic-data generation and fine-tuning pipelines, running today to produce privacy-safe corpora at scale.
Cosmos
Foundation-model-style generation of synthetic knowledge scenarios as the engine deepens.
Omniverse
Simulation and visualization of the organizational digital twin, with custom CUDA simulation kernels.
Large-scale generation and simulation run on H100, DGX, and A100, on NVIDIA AI Enterprise for regulated synthetic-data workloads. Cosmos and Omniverse capabilities are aspirational and deepen over time.
Defensibility
Data from your graph, not a generic set.
Synthetic data generated from a customer's own trained graph is far more useful than off-the-shelf synthetic data, and the digital twin is unique to that org's knowledge. Quality compounds with the graph and the flywheel behind it.
Where it sits
The engine that improves the rest.
Synthos simulates the Graphos graph, feeds training data to Mnemo and Nexora, and supplies edge cases to Verix. It stays inside the knowledge domain as the simulation and data layer.