Developer transparency

FinanceGPT ML Service

Architecture, model boundaries and verified runtime posture for FinanceGPT learned-model inference and evaluation.

Runtime status
Operational
Last verified 2026-08-12T18:38:17Z
Numerical runtime
NumPy 1.26.4
Numerical imports must pass before the service is marked ready.
Model runtime
PyTorch 2.13.0+cu130
CPU execution detected.
Inference
Smoke test passed
Operational status requires a deterministic service and inference verification.

Architecture

Laravel remains the control plane. The private FastAPI service receives bounded numeric payloads and returns governed model outputs.
FinanceGPT API / Quant workflows
Governed ML gateway
Model registry + version policy
Private FastAPI ML service
Evidence, evaluation and model/version lineage

Readiness gates

Each gate is independently verified; installed packages alone do not establish runtime readiness.
NumPy numerical importPassed
PyTorch importPassed
FastAPI service startup / healthPassed
Model inference smoke testPassed
Execution deviceCPU

Capabilities

Learned models complement deterministic FinanceGPT quantitative models; availability depends on the verified runtime and model approval state.
Financial anomaly modelsReconstruction-based anomaly signals with governed evaluation and model cards.
Conditional financial scenariosBounded generative scenario samples with explicit regime and stress inputs.
Time-series path generationVersioned sequence generation with deterministic seeds and evidence lineage.
Model evaluation & driftCandidate evaluation, human activation and drift review rather than silent promotion.

Operating boundary

No database credentials

The ML service is not given FinanceGPT database credentials and cannot independently fetch customer records.

Derived numeric inputs

Training and inference payloads are bounded to approved numeric feature schemas rather than raw document or transaction text.

No execution authority

Learned-model outputs cannot place trades, make payments, approve actions or bypass the governed execution layers.

Hardware is a separate fact. CUDA runtime packages do not prove a physical NVIDIA GPU exists. CPU/GPU posture is reported only from verified runtime evidence.

Verification outputs

Public status intentionally excludes internal host names, filesystem paths, credentials and deployment commands.
status: operational
checked_at: 2026-08-12T18:38:17Z
python: 3.12.13
execution_device: cpu