Technical overview
An overview of how AscentML is built, how it handles customer data, and where its current boundaries are.
This page is written for architects, engineers, and security reviewers evaluating the platform. Detailed documentation is available to customers inside the platform and during vendor review.
-
1
Architecture
AscentML is a hosted, multi-tenant SaaS platform. The customer platform is separated from the internal environment where AscentML develops and trains models.
Customer-facing services
- Web application — projects, datasets, model review, deployments, batch predictions, administration
- Worker — validation, screening, batch jobs, notifications, and other background work
- Inference runtime — serves published models for API and batch predictions
Internal boundary
Customers submit training requests. They do not run training. A dedicated integration layer carries requests to the internal training environment and returns published models. That layer can only reach the training-request and publication surface—nothing else in the customer platform.
-
2
Tenant isolation
- Every customer-owned record carries a customer identifier
- Tenant predicates are required on every customer-facing query
- Separate least-privilege database roles for the web/worker tier, the inference runtime, migrations, and internal integration
- Tenant-prefixed object storage, with signed URLs scoped to one object, one action, and a short expiration
- Tenant-scoped cache keys
- Automated tests for cross-tenant leakage
-
3
Data handling
CSV and Parquet files are uploaded directly to object storage through short-lived presigned URLs. Storage keys are derived by the server, and uploads are verified for existence, size, and checksum before any state change.
Datasets are immutable and versioned. Each version is profiled for null rates, constant and near-constant columns, and leakage, then screened for sensitive information. High-confidence findings are rejected; ambiguous findings are quarantined pending customer and AscentML approval.
Derived features are expressed in a restricted, allowlisted expression grammar. Expressions are parsed into a safe syntax tree—never evaluated as code—and produce identical results in validation and in serving.
-
4
Model lifecycle
Training requests follow an explicit state machine in which each transition is owned by either the customer or AscentML. Customers can follow a request through review, training, validation, and publication.
A published model version is an immutable bundle: a versioned manifest describing inputs, imputation, derived features, and model stages, plus checksummed artifacts. Models are served with ONNX Runtime. The runtime verifies checksums and manifest schema before loading a model.
Input rows that cannot be processed return explicit per-row errors. Rows are never silently dropped.
-
5
Deployment and inference
Deployments can split traffic across model versions using integer percentages that total 100. Each allocation is stored as an immutable revision.
Routing is deterministic. A caller-supplied routing key is hashed with the revision, so the same caller receives the same model version for the life of that revision. Concurrent allocation changes are protected by optimistic locking.
Predictions are available through a versioned REST API or guided batch jobs. Per-prediction explanations are computed on demand against the original model version.
-
6
Reliability
- Durable, database-backed job queue with heartbeats, bounded retries, and recovery of stalled jobs
- Transactional outbox, so business changes and their events commit together
- Idempotency keys for batch submission, training requests, publication, deployment changes, and upload completion
- Correlation IDs carried from the API through jobs, inference, logs, and audit
- Automated database backups, point-in-time recovery, and versioned object storage
Committed customer data is designed to survive service restarts and failures. Availability targets and recovery objectives are shared during vendor review.
-
7
Current scope
We would rather be clear about what is not included than leave reviewers to discover it.
Not part of the current release
- Self-service training, AutoML, or fine-tuning
- Notebooks or SQL workbenches
- Single sign-on (SSO, SAML, OIDC)
- Drift monitoring or automated retraining
- Rules-based routing or automatic promotion and rollback
- Multi-dataset joins or feature stores
- Multi-region or customer-managed infrastructure
- Regulated workloads such as HIPAA or FedRAMP
Several of these, including SSO, drift detection, and multi-zone failover, are on the roadmap. We will tell you which are planned and which are not.
Need more detail for your review?
We can provide deeper architecture information and answer specific questions from your engineering, security, and IT teams.