⚓ Keel

Agentic AI for B2B EDI supply chains - it watches your trading partners, flags the revenue at risk when an order fails, and recommends an alternative supplier before the gap becomes a problem.

How it works · Quickstart · What it demonstrates

---

The problem

In a B2B supply chain, orders flow between trading partners as EDI messages (X12 / EDIFACT - an 850 purchase order, an 855 acknowledgement, an 856 shipping notice, and so on). When one of those orders silently fails - a partner endpoint goes down, an 850 is never acknowledged, a message is rejected - the business often doesn't find out until a shipment is late and revenue has already slipped. The information needed to react (which order, which partner, what it's worth, who else could supply it) is scattered across integration platforms, ERP systems, and partner records.

Keel closes that gap. It continuously watches EDI partner and order activity, quantifies the revenue impact of a failure the moment it's detected, and uses an agentic loop to recommend a concrete next action - including which alternative supplier to onboard or route business to.

What it does

• 🔍 Tracks B2B EDI trading partners - monitors the health and message flow of partners across the supply chain, so a stalled or failing partner surfaces immediately instead of days later.
• 💸 Quantifies revenue at risk - when a purchase order fails or stalls, Keel extracts the order details and calculates the revenue impact, turning a silent integration error into a ranked, business-prioritized alert.
• 🔌 Integrates through your API & integration layer - connects to existing Integration / API Management platforms to pull live order and partner data rather than duplicating it, fitting into the architecture you already run.
• 🤝 Recommends an alternative supplier - reasons over partner capability, history, and the at-risk order to suggest the best alternative supplier to onboard or do business with, and can kick off the onboarding workflow.

🎬 Demo

---

🏗️ How it works

flowchart TB
    subgraph Partners["B2B EDI Trading Partners"]
        P1[Supplier A]
        P2[Supplier B]
        P3[Supplier C]
    end

    Partners -->|"X12 / EDIFACT<br/>850 · 855 · 856 · 810"| INT["Integration &amp; API Management<br/>(your iPaaS / APIM layer)"]
    INT --> CORE

    subgraph CORE["Keel · Agentic Core"]
        direction TB
        MON["Partner &amp; Order Monitor<br/>detects failed / stalled POs"]
        RISK["Revenue Impact Analyzer<br/>extracts order, scores $ at risk"]
        REC["Alternative-Supplier Agent<br/>plan → evaluate → recommend"]
        MON --> RISK --> REC
    end

    RISK --> DASH["Revenue-at-Risk Dashboard"]
    REC --> ALERT["Prioritized Alerts"]
    REC -->|"initiate"| ONB["Supplier Onboarding Workflow"]

Loading

The flow in words:

1. Monitor — Keel reads partner and order activity through the integration/API layer and detects when a purchase order fails, is rejected, or goes unacknowledged.
2. Quantify — it extracts the failed order's details and computes the revenue at risk, so failures are ranked by business impact, not just timestamp.
3. Reason — the agentic core plans over partner capabilities and order requirements, evaluates candidate suppliers, and reflects on whether a recommendation is well-supported.
4. Act — it surfaces a prioritized recommendation and can initiate onboarding for the alternative supplier.

Key design decisions

Decision	Why
<e.g. "Read order data via the existing API Management layer instead of a new ERP connector">	<e.g. "Avoided duplicating a system of record and kept the integration footprint small.">
<e.g. "Score revenue-at-risk before recommending">	<e.g. "Lets the team triage by business impact; a $2M order outranks a $5k one.">
<e.g. "Agent abstains when partner data is incomplete">	<e.g. "A bad supplier recommendation is costlier than asking a human to confirm.">

---

⚡ Quickstart

# 1. Clone
git clone https://github.com/algoshank-pat/<REPO>.git
cd <REPO>

# 2. Install
python -m venv .venv && source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install -r requirements.txt

# 3. Configure
cp .env.example .env
# Add your integration/API-management endpoint + credentials and LLM key to .env

# 4. Run
python -m keel watch                 # start monitoring partner/order activity
python -m keel simulate-failure 850  # demo: simulate a failed PO and see the recommendation

See docs/ for connecting to your integration platform and configuring the revenue model.

---

📈 What this demonstrates

This project demonstrates an end-to-end pattern for AI-assisted supply-chain resilience:

• Designing an agentic system that sits on top of existing integration/API-management infrastructure rather than replacing it.
• Translating a low-level integration failure (a failed EDI message) into a business signal (revenue at risk).
• Closing the loop from detection → impact → recommendation → action.

If you've run it, add your real figures:

Metric	Result
Time-to-detection of a failed PO	<e.g. "minutes vs. hours manually">
Orders monitored across partners	<e.g. "N partners, M orders/day">
Alternative-supplier suggestion accuracy	<e.g. "evaluated on K historical cases">

---

🗂️ Project structure

<REPO>/
├── src/keel/
│   ├── monitor/        # EDI partner & order monitoring
│   ├── risk/           # revenue-impact analysis
│   ├── agent/          # alternative-supplier reasoning loop
│   ├── integration/    # connectors to API / integration management
│   └── api/            # service + dashboard endpoints
├── docs/               # architecture & integration setup
├── tests/
└── assets/             # demo gif, diagrams

🧰 Built with

Python 3.11 · <your agent framework — e.g. LangGraph / Semantic Kernel> · <LLM — e.g. Azure OpenAI / OpenAI> · <integration platform — e.g. Azure API Management / MuleSoft / Boomi> · <EDI tooling, if any>

🗺️ Roadmap

• <e.g. "Auto-draft the EDI onboarding pack for a recommended supplier">
• <e.g. "Confidence scoring on revenue-impact estimates">
• <e.g. "Support for additional EDI document types">

🤝 Contributing

See CONTRIBUTING.md. Open an issue before large changes.

📄 License

MIT — see LICENSE.

---

_{Built by Shashank Patel · LinkedIn · Email}