How GenericML works

GenericML turns ML experimentation into a configuration problem rather than a re‑engineering exercise: add domain features, pick a label, generate candidate models, and keep everything traceable.

The flow

Model the domain (DDD)

Define bounded contexts + aggregates so semantics are stable and leakage is reduced.
Define a decision problem

Make the decision explicit (what action, what constraints, what outcomes).
Create a VectorSchema

Versioned features (float / float[]), units, transforms, windowing rules, leakage controls.
Generate vectors

Produce VectorInstances from domain events/telemetry. Store big arrays externally; keep hashes/refs in the graph.
Train ModelPacks (AutoML)

CPU‑first training + validation, producing deployable artefacts with evidence and validity envelopes.
Ensemble the signals

Weighted, stacked, rule+model, confidence‑aware, or regime‑gated (mixture‑of‑experts).
Ship into apps/agents

Expose via REST/gRPC/serverless; decisions happen at point‑of‑use, not in a separate notebook pipeline.
Log evidence + outcomes in Neo4j

Trace everything: data → vector → model → ensemble → evidence → decision → outcome.
HITL continuous improvement

Experts review errors, adjust thresholds, approve updates, and keep models aligned with reality.

One practical rule

If you can’t explain a prediction in terms of the domain model, you probably don’t have the right bounded context or vector schema yet.