Back to selected work

ML systems / decision intelligence

SentinelFlow

Synthetic-data reference implementation

A transparent fraud-risk workflow demo built to make modelling, policy, and review surfaces inspectable.

What it is

A transparent fraud-risk workflow demo built to make modelling, policy, and review surfaces inspectable.

What it demonstrates

SentinelFlow joins event-time features, a trained model, calibrated policy routing, and an operator-facing dashboard while keeping the synthetic-demo boundary explicit.

My role

Independent builder

Project status

Public synthetic demo

01The problem

Risk workflows can hide important modelling and policy choices behind a single score. Operators need the score, its operating context, and a clear boundary between inference and action.

02System architecture

SYSTEM MAP

Next.js dashboard
FastAPI typed API
Redis event-time state
Model inference and policy layer
PostgreSQL production path / SQLite tests
MLflow-tracked artifacts

03Model and product pipeline

FLOW

Synthetic events
Feature construction
Temporal split
Training and calibration
Policy routing
Review and explanation

04Engineering decisions

01

Separate prediction from policy

A calibrated probability is not itself a business action; routing remains visible and independently adjustable.

02

Use event-time state

Redis supports bounded stateful signals without pretending the demo is a live banking integration.

03

Track training artifacts

MLflow keeps model runs and related artifacts inspectable.

04

Design for review

Dashboard language frames outputs as decision support rather than autonomous fraud decisions.

05Product screenshots

SentinelFlow risk operations dashboard
Live public demo: operations dashboard and decision-policy surface.
SentinelFlow live fraud simulator
Live public demo: the FastAPI-backed scoring and decision-policy simulator.

06Results and evaluation

  • - The repository documents synthetic generation, temporal splitting, calibration, and reproducible artifacts.
  • - This page intentionally does not present a production fraud-detection claim or live-bank performance metric.

07Limitations

  • - Synthetic data only; not a banking, compliance, or production-risk system.
  • - Counterfactual signals are non-causal and require human interpretation.

08What I learned

A score needs policy context to be operationally useful.

Temporal evaluation matters when events have order.

Clear demo boundaries build more trust than inflated production claims.

FastAPINext.jsRedisPostgreSQLMLflow