Scenarios
Fintech x Agentic AI x JacHacks 2026
Fraud investigations with real agent behavior, not just a score.
PayGuard Agent reviews suspicious transactions and returns ALLOW, ESCALATE, or BLOCK. Every result includes a transparent trace: plan, tool calls, memory context, and final rationale.
Tool Checks
Decisions
Paid APIs
Core Capabilities
Planning
Builds a deterministic investigation plan per transaction for consistent demos.
Tool Use
Executes user profile, velocity, merchant risk, amount outlier, and location checks.
Memory
Stores prior cases locally and reuses them in future decisions.
Multi-Step Reasoning
Combines signals into a weighted score and chooses a final action.
Explainability
Outputs a full trace from plan to decision with human-readable reasons.
Hackathon Friendly
No paid API required. Fully local, reproducible, and demo-ready.
Agent Workflow
- Plan checks: choose validations based on transaction context.
- Call tools: fetch profile, velocity, merchant risk, and anomaly signals.
- Read memory: compute prior fraud rate for the user.
- Score: convert observations into a risk score from 0 to 100.
- Decide: return ALLOW, ESCALATE, or BLOCK with top reasons.
Each step is logged so judges can verify autonomy, tool usage, and rationale.
Demo Flow
1. Install in one command
Judges run a single bootstrap command from GitHub and get a ready environment fast.
2. Launch deterministic run
Use reset-memory mode to keep scenario outcomes stable during live judging.
3. Review the agent trace
Output shows PLAN -> TOOL -> MEMORY -> DECISION for transparent reasoning.
4. Validate persistence
The memory file stores previous investigations and influences later risk posture.
How To Run
Install once from GitHub, then run the demo locally.
Install once (Windows PowerShell)
powershell -ExecutionPolicy Bypass -Command "iwr https://raw.githubusercontent.com/Ranojitdas/Finance-app-jac/main/scripts/bootstrap_windows.ps1 -UseBasicParsing | iex"
Install once (Linux / macOS)
curl -fsSL https://raw.githubusercontent.com/Ranojitdas/Finance-app-jac/main/scripts/bootstrap_unix.sh | bash
Standard run
python src/app.py
Deterministic run (recommended)
python src/app.py --reset-memory
Windows fallback (if python is missing in PATH)
py src/app.py --reset-memory
Run with virtual environment Python (best for judges)
.venv/bin/python src/app.py --reset-memory (Linux/macOS) | .\\.venv\\Scripts\\python.exe src/app.py --reset-memory (Windows)
Tip: Use reset mode before recording your 3-minute demo for stable output.
Output appears in terminal with full investigation traces for every scenario.
Judge tip: run from project root so data and memory paths resolve correctly.
Included Scenario Set
Baseline legitimate purchase
Account takeover cross-border
New user ambiguous escalation
False positive travel-like pattern
Mule pattern rapid small transactions
Demo Results (Explainability)
Loaded from a local demo JSON to showcase the agent explainability labels and summaries.
Key Project Paths
- Runner: src/app.py
- Agent logic: src/agent.py
- Jac logic: jac/payguard_agent.jac
- Scenarios: data/scenarios.json
- Memory: memory/cases.json
Ready To Run?
Install in one command, run the deterministic demo, and show the trace in your submission video.
FAQ
Do I need external APIs?
No. The current project is fully local and works without external APIs.
Is Jac included in this project?
Yes. Jac logic is in jac/payguard_agent.jac. Python currently runs the local demo path. You can wire Jaseci runtime for Jac-primary execution in final submission.
Can I host this page quickly?
Yes. Run a local static server and open web/index.html in your browser.