PyGuard
Beta release

Turn raw fraud flags into actionable intelligence.

PyGuard is an AI fraud analyst that reads your flagged transaction table, writes an executive-ready report, and keeps answering follow-up questions inside Slack. It does not replace your team. It takes the mechanical review work off their plate so the decisions still belong to humans.

Runs inside
Slack
Output
PDF reports
Human review
Always
PyGuard
#fraud-ops • just now

February 2025 fraud intelligence report attached below. 2,189 transactions processed. 788 flagged (36.0%). Two attack waves identified on Feb 10-14 and Feb 18-20. Full report follows as PDF.

fraud_report_20260420.pdf
9 pages • generated just now
Reply in threadWhy did you rate chargeback abuse as moderate and not critical?
Promotional video

Ninety seconds inside a real Slack workspace

A short walk through of the live product, from a flagged transaction table to an in-thread report, a follow-up question, and a LaTeX edit. No slides and no mockups.

A short, 1-3 minute promotional video showing the live PyGuard product in action: a Slack user asks for a fraud intelligence report, PyGuard posts a PDF into the thread, and then fields a follow-up question and a LaTeX edit inside that same thread. No slides and no mockups.

Live
Promo1-3 min

PyGuard in sixty to ninety seconds

Live product flow inside a real Slack workspace.

Live
Demo6-7 min

End-to-end walkthrough

Installation, first report, follow-up Q&A, and LaTeX edits. Designed so a new user can reach the primary workflow without reading anything else.

Demo video

Follow the workflow yourself

Pairs with the setup guide. If you watch this once, you can run the end-to-end workflow on your own Slack workspace without referring back to the docs.

An end-to-end walkthrough covering installation, first report generation, in-thread Q&A, and LaTeX edits. Designed so a first-time user can follow along and reach the primary workflow without reading any other documentation.

The problem we see every week

Flag fatigue is the real bottleneck for SMEs

Most teams already run some form of fraud detection. The hard part is never the detection itself, it is what happens after the model finishes and hands you a long table of flagged transactions.

Reality check
Thousands

Flagged rows per month on a typical SME fraud feed.

Reality check
No team

Very few SMEs can justify a full-time fraud analyst or data scientist.

Reality check
Hours

Spent every week pivoting spreadsheets and writing summary emails.

Executives and operations leads end up as accidental fraud analysts. They spend the afternoon inside a spreadsheet instead of working on the business. We built PyGuard because we wanted to soften that specific manual step, not remove the human from the loop.

Our fresh perspective

The novelty is not the analysis. It is the action.

Plenty of tools can describe fraud. What very few of them do is translate that description into prioritised, specific actions a small team can hand straight to operations, engineering, or risk.

Narrative over spreadsheets

PyGuard writes the report the way a human analyst would: named patterns, business framing, clear numbers, and the parts of the month that actually changed.

Actions you can run tomorrow

Every report ends with prioritised recommendations, split into IMMEDIATE, SHORT-TERM, and MEDIUM-TERM. Each one is specific enough to assign.

Answers inside the thread

Once the PDF is in Slack, PyGuard stays in the thread. Ask a follow-up, request a rewrite, or tighten the executive summary without leaving the conversation.

How it works

From CSV to decision in four steps

The goal is to keep the conversation in one place. You never have to leave Slack, and you never have to open a spreadsheet to understand what PyGuard did.

    01

    Raw transaction table

    Your detection model produces the usual long CSV of flagged transactions. PyGuard reads it as input, not as the final answer.
    02

    Ask in Slack

    A teammate sends one message asking for a fraud intelligence report. No new dashboard, no new login, no attachment needed.
    03

    PDF lands in thread

    PyGuard posts a short summary into the channel and uploads a formatted PDF report into the thread, with named patterns and prioritised recommendations.
    04

    Review, ask, edit

    Reply in the same thread with questions or edits. PyGuard is the expert on its own report and ships a new version on request.
Architecture

A thin agent layer on top of your existing detection

PyGuard is not a detection engine. It sits above your model, reads the flagged data, and orchestrates a small set of sub-agents to produce a report, answer questions, and apply edits in place.

flowchart LR
  User["SME operator in Slack"] --> Slack["Slack Socket Mode"]
  Slack --> API["PyGuard backend"]
  API --> Memory["Memory agent"]
  Memory --> Fraud["Fraud orchestrator"]
  Fraud --> Data["Fraud data agent"]
  Fraud --> Pattern["Pattern agent"]
  Fraud --> Report["Report agent"]
  Report --> PDF["PDF report in thread"]
  PDF --> User
A simplified view of the request path. See the full architecture page for every component and data flow.
Human in the loop

We reduce manual review. We do not remove the human.

PyGuard is built to make an accidental fraud analyst faster at the mechanical parts of their job. Every decision still belongs to a person, and every report still gets reviewed.

  • Reports are a starting point for human review, not a final decision.
  • Every actionable recommendation is owned by someone on your team before it ships.
  • Edits are explicit requests. PyGuard never rewrites a report on its own.
  • Every previous version is saved locally so a rollback is one step.

PyGuard is not trying to replace a fraud analyst. It is trying to take the most tedious parts of their job and hand them a better starting point.

Our honest positioning

Detection is already solved for most SMEs. What is not solved is turning the detection output into a decision a small team can act on this week.

What we heard from operators