The Challenge
PropCFlow Inc. processes more than 500,000 transactions a day. Their fraud detection worked — eventually. Transactions were scored in batch, and the batch took 8 hours.
In fraud, latency is not a performance metric. It is the fraudster’s head start. An 8-hour scoring window meant that by the time PropCFlow’s models flagged a transaction, the money had often moved twice more. Fraudulent activity that a real-time system would have blocked at the point of transaction instead became a recovery case — slower, costlier, and frequently unsuccessful.
The batch architecture created three compounding problems:
- Detection arrived after the damage. Flags fired on transactions that had already settled. The fraud team was doing archaeology, not prevention.
- The models were opaque. When a score did arrive, analysts and compliance teams could not see why a transaction was flagged, making investigations slow and regulator conversations uncomfortable.
- Reporting was manual. Leadership visibility into fraud exposure depended on hand-built reports assembled from the previous day’s batches.
PropCFlow’s mandate to OmniMinds was blunt: score every transaction before it completes, explain every score, and stop losing money to the clock.
Our Solution
We rebuilt the fraud pipeline as a streaming system, end to end.
Kafka + Delta Lake streaming backbone. Transactions now flow through Kafka the moment they occur, landing in Delta Lake with full ACID guarantees. The same stream serves real-time scoring and historical analytics, so the online models and the training data can never drift apart — one source of truth, milliseconds fresh.
XGBoost models with SHAP explainability. We trained gradient-boosted models on PropCFlow’s transaction history and put them directly in the streaming path. Every score ships with SHAP attributions: the exact features that drove the decision, in plain terms an analyst — or a regulator — can read. “Flagged” became “flagged because velocity anomaly + device mismatch + amount deviation,” which cut investigation time and made model governance defensible.
MLOps on Spark and Airflow. Retraining, backtesting, and deployment run as orchestrated Airflow pipelines on Spark. Models retrain on fresh fraud patterns on schedule, pass evaluation gates against holdout sets, and promote automatically — with rollback if precision degrades. Fraud evolves weekly; a model frozen at launch would decay in months. This one does not.
Automated Power BI reporting. Fraud exposure, model performance, and recovery metrics flow into Power BI dashboards automatically. The manual reporting cycle is gone; leadership sees the fraud posture of the business in near real time.
The OmniMinds pod pairing senior data engineers and ML engineers with AI agents handled schema mapping, pipeline testing, and evaluation harness work at machine speed — one of the reasons a system of this scope shipped on a services timeline rather than an internal-platform-team timeline.
The Results
- 47 milliseconds of fraud-scoring latency, down from 8 hours. That is a ~600,000× improvement. Every one of PropCFlow’s 500K+ daily transactions is now scored while it is still in flight — fraud is blocked, not investigated.
- 92% precision. High precision means the fraud team’s queue is full of real fraud, not false positives. Analysts investigate signal, and legitimate customers stop getting blocked for looking suspicious.
- $50M recovered. The headline number. Real-time detection plus explainable scores turned fraud losses into recovered funds at a scale that repaid the engagement many times over in its first year.
There is also a quieter result: compliance conversations changed. With SHAP explanations attached to every decision and full lineage in Delta Lake, PropCFlow can show regulators exactly how and why its models decide — a posture very few fintechs at this scale can claim.
Why It Worked
The engagement was priced and scoped the way OmniMinds always works: against outcomes. Latency under a defined threshold, precision above a defined bar, reporting automated. Not “a team of eight for twelve months” — a result, delivered.
Getting from 8 hours to 47 milliseconds is an architecture problem before it is a modeling problem, and that is where senior talent is non-negotiable. The engineers who designed this pipeline carry backgrounds from IBM, TCS, and DXC, and it shows in the decisions: one streaming backbone serving both scoring and training, explainability as a first-class requirement rather than a retrofit, and MLOps gates that assume models decay because they do.
The AI-augmented pod structure compressed everything around those decisions — test coverage, pipeline scaffolding, data validation — so the humans spent their hours on the choices machines cannot make. That is the OmniMinds model: AI agents and senior engineers as one team, accountable for a number. In PropCFlow’s case, several numbers — 47 ms, 92%, and $50M.