Agent 01 · Fintech
FraudShield Agent
Explains and triages risky card transactions for analysts.
Narrow scope firstCoreLangGraph
One offline run on demo data with planted problems (recorded platform test, November 2026).
- Client problem
- Fraud analysts drown in alerts; bad transaction data (blank or negative amounts, duplicates) hides real fraud.
- What it does
- Cleans the transaction feed, scores every transaction with a trained gradient-boosting model (plus an IsolationForest anomaly score), and writes a plain-English reason for each high-risk one.
- How it heals
- Quarantines broken rows, removes duplicates, auto-closes low-risk alerts; high-risk holds go to a human in the Control Tower.
- Human gate
- Blocking a card or holding money always needs analyst approval.
- Typical buyer
- Payment companies, NBFCs, card issuers (risk/fraud ops lead)
- Built with
- LangGraph self-healing loop
- Demo data
- Kaggle mlg-ulb/creditcardfraud (284,807 European card transactions).
- Tables
- transactions_cc ← Kaggle mlg-ulb/creditcardfraud
- Target outcome
- Target: 40-60% less alert-review time; 100% of data faults caught on the demo set.
- Path to production
- Keep real-time decisions in the client's rules engine; DigiQuench adds explanation, triage and data-quality healing.
| Issue | Risk | Proposed fix | Count |
|---|---|---|---|
| missing_amount 'amount' is empty | low | Quarantine and request re-send from the switch | 5 |
| out_of_range_amount amount=-189.47 outside [0, None] | low | Correct sign or reverse the faulty entry | 4 |
| duplicate_record Duplicate on txn_id | low | Drop duplicate transaction event | 5 |
| high_fraud_risk Fraud score 0.99 (anomaly 0.54) on amount 59.83 | high | Hold transaction and open analyst case | 25 |
How we demo it
- Show 20,000 transactions with injected blanks and duplicates.
- Run FraudShield: data repaired, model scores, 25 high-risk holds appear in Approvals.
- Approve two holds live and show the audit trail.
Built on AWS
Runs in the platform container on EC2 Graviton (or on demand in Lambda), writes healed tables to the S3 lakehouse, uses Amazon Bedrock only to write explanations, and reports six metrics to CloudWatch.
KYC AutoFlow
Checks KYC records, catches bad documents and duplicates before they reach underwriting.
03 · FintechAML Monitor
Watches mobile-money flows for structuring and round-tripping, and drafts case files.
04 · FintechLoanProcess
Repairs loan applications and flags high-risk ones with a model score and reasons.
Step into DigiQuench.
Agents that heal your data.
Watch an agent find planted problems, fix the safe ones and wait for your approval on the rest. Thirty minutes, live.