Shurokkha সুরক্ষা
Demo prototype: all customers, wallets and transactions are synthetic. No real personal data.

Track 01 Trust & Risk · Track 03 Financial independence

Stop the scam before the money leaves.

Shurokkha scores every transfer in milliseconds, explains risk to the customer in plain Bangla, finds money-mule rings in the transaction graph and gives analysts an AI case summary grounded only in evidence. It answers three questions: What happened? Why is it risky? What should upay do next?

Try the customer demoOpen analyst consoleSee impact metrics →

Rahima's story

Persona · remittance receiver, first smartphone

  1. 1. A caller says she won a prize and must pay a 2,000 BDT fee.
  2. 2. She types a number she has never paid. Shurokkha sees 30+ new senders and fast cash-outs on it.
  3. “একটু থামুন, এটি প্রতারণা হতে পারে”
  4. 3. She cancels. The wallet goes to an analyst with the full mule-ring picture.
Loading…

How a decision is made

Input → Intelligence → Action → Feedback (guideline §12)

  1. 1. Input
    Transfer request
    Customer app / partner API calls /score before money moves.
  2. 2. Features
    Point-in-time features
    Velocity, novelty, device, SIM-swap, fan-in, graph community.
  3. 3. Detectors
    4 plug-in detectors
    Rules · Isolation Forest · LightGBM · mule-graph risk.
  4. 4. Policy
    YAML policy engine
    Business rules turn calibrated scores into ALLOW / WARN / HOLD.
  5. 5. Explain
    SHAP reason codes
    Only true reasons, rendered in Bangla and English.
  6. 6. Action
    Customer & analyst
    Scam interrupt, human review, AI case summary grounded in evidence.
  7. 7. Feedback
    Labels → retrain
    Analyst labels feed the model registry and threshold tuning.

Responsible by design

  • • 100% synthetic data, no real PII
  • • The LLM narrates; it never decides
  • • No permanent auto-block: HOLD goes to a human
  • • Fairness audit across division, age, KYC, tenure

Built to adapt

  • • New detector = one file + one YAML line
  • • Policy thresholds hot-reload without redeploy
  • • Versioned model registry with one-click activate
  • • Feature flags for every panel

Built to survive

  • • LLM down → deterministic template summaries
  • • Model missing → rules-only mode, flagged degraded
  • • Idempotent scoring, audit log, health probes
  • • Redis optional, in-memory fallback