State conflicts
A transaction record can disagree with the events recorded around it.
Live Experimental Prototype
PayResolve
Payment failures shouldn't require guesswork. PayResolve is an experimental payment operations intelligence platform for investigating transaction failures, anomalies and reconciliation exceptions.
PAYMENT
INTELLIGENCE
TRANSACTION
RECONCILIATION
INCIDENT
AI INSIGHT
INTELLIGENCE
01 / The Problem
When a payment has no clear story.
Payment transactions can move through several states and systems. When those records disagree, operations teams may have to manually piece together what happened before deciding what needs attention.
Reconciliation gaps
Missing settlements, reversals or events can make transaction outcomes unclear.
Investigation overhead
Teams need evidence and context before they can determine what requires attention.
02 / Product Concept
Turn transaction data into an investigation workflow.
PayResolve combines deterministic reconciliation rules, incident classification, anomaly detection and an AI explanation layer into one operational workflow.
01 · Transaction Data
Records and event history form the evidence base.
02 · Reconcile
Rules identify mismatches and exceptions.
03 · Classify
Exceptions become operational incident categories.
04 · Detect
Anomaly detection surfaces unusual records.
05 · Explain
AI summarizes evidence and what remains unknown.
03 / What I Built
From product definition to working prototype.
I designed the product workflow, generated synthetic payment data, built the investigation logic and turned the outputs into an interactive operations interface.
Reconciliation Engine
Checks transaction and event records for missing or conflicting states, settlements, reversals and other exceptions.
Incident Classification
Maps detected exceptions into operational categories and investigation priorities.
Anomaly Detection
Uses Isolation Forest to identify transaction records that differ from expected patterns.
AI Investigation
Uses the evidence layer to generate a structured explanation without making autonomous financial decisions.
Operations Dashboard
Provides overview metrics, incident analysis, anomaly analysis, transaction investigation and timelines.
API Layer
FastAPI endpoints expose summaries, transactions, incidents, anomalies, timelines and AI investigations.
04 / Prototype Data
The prototype is grounded in synthetic evidence.
The test dataset contains intentionally injected operational scenarios so the investigation workflow can be evaluated without using real customer or financial data.
10,020
Test transactions
Generated payment records used to test the
investigation workflow.
1,524
Reconciliation exceptions
Records identified by the deterministic
reconciliation layer.
501
Detected anomalies
Records flagged by the experimental anomaly
detection model.
Injected scenarios
Missing settlements · Status conflicts · Missing events · Provider timeouts · Potential duplicate payments · Debit without settlement.
Important boundary
The dataset is synthetic. PayResolve does not process real money, connect to real banks or investigate real customer transactions.
05 / AI Investigation
AI explains the evidence. It doesn't invent the story.
The deterministic engine establishes the facts first. The AI layer then uses that evidence to structure an investigation summary and identify what remains unknown.
Evidence First
The investigation prompt is deliberately constrained to the available transaction, event, reconciliation and anomaly evidence.
01
Investigation Summary
Summarizes the transaction outcome using the
available evidence.
02
Evidence
Surfaces the recorded transaction, event and
classification evidence.
03
What Is Unknown
Explicitly identifies information that the evidence
does not establish.
04
Recommended Next Step
Suggests operational investigation steps rather
than autonomous financial actions.
06 / Product Decisions
Designed around operational trust.
Evidence First
Facts are established before AI is asked to explain them.
Human in the Loop
The system supports investigation rather than replacing operational judgment.
Traceability
Investigations remain connected to the transaction and event evidence behind them.
Privacy by Design
The prototype uses synthetic data and avoids real customer payment information.
Operational Reliability
Exceptions are categorized so teams can understand what requires attention and who may own it.
Explainable Experimentation
Experimental models are presented as signals for investigation, not unquestionable conclusions.
07 / Technical Stack
Built as a product experiment.
The prototype combines data processing, machine learning, API design, an interactive interface and a constrained generative AI layer.
Python
Pandas
NumPy
Scikit-learn
Streamlit
FastAPI
Groq
Llama 3.3
CSV Data Pipeline
08 / Limitations
A prototype, not production financial infrastructure.
No real customer, bank or payment-provider data is used.
The anomaly detector demonstrates an approach and requires further validation before production use.
AI-generated explanations are decision-support outputs and require human review.
See the investigation workflow in action.
Explore the live prototype or return to the Fintech & Payments Lab to see the broader product thinking behind PayResolve.