2nd/10_Wiki/Topics/AI_and_ML/Supply-Chain.md

---
id: wiki-2026-0508-supply-chain
title: Supply Chain
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [SCM, Supply Chain Management, 공급망]
duplicate_of: none
source_trust_level: A
confidence_score: 0.9
verification_status: applied
tags: [supply-chain, logistics, security, sbom, ai-optimization]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: python
  framework: ortools
---

# Supply Chain

## 매 한 줄
> **"매 supply chain 의 end-to-end network — raw material부터 end customer까지의 flow"**. 매 2026 supply chain 의 두 축: AI-driven optimization (demand forecast, route, inventory) 와 security (SBOM, supply chain attack defense). 매 SolarWinds·xz-utils 사건 이후 software supply chain 의 first-class 보안 concern.

## 매 핵심

### 매 5 components
- **Plan**: demand forecast, capacity planning, S&OP.
- **Source**: supplier selection, contract, procurement.
- **Make**: production, quality, scheduling.
- **Deliver**: warehousing, transportation, last-mile.
- **Return**: reverse logistics, recycling, RMA.

### 매 AI 적용 영역
- **Demand forecasting**: Transformer-based time series (TimesFM, Chronos), LSTM 의 retire.
- **Route optimization**: OR-Tools VRP + RL hybrid.
- **Inventory**: (s,S) policy + safety stock dynamic adjustment.
- **Anomaly detection**: shipment delay prediction, fraud.
- **Supplier risk**: graph neural network on supplier dependency graph.

### 매 software supply chain security
- **SBOM** (Software Bill of Materials): SPDX, CycloneDX format.
- **Sigstore**: keyless signing, transparency log.
- **SLSA** (Supply-chain Levels for Software Artifacts): level 1-4 framework.
- **Attack surface**: dependency confusion, typosquatting, malicious maintainer.

### 매 응용
1. **E-commerce**: Amazon FBA — AI demand forecast → DC pre-positioning.
2. **Manufacturing**: Toyota JIT 의 AI evolve — predictive lead time.
3. **Software security**: GitHub Dependabot + Sigstore + SLSA Level 3.

## 💻 패턴

### 1. Demand forecast (Chronos)
```python
from chronos import ChronosPipeline
import torch
import pandas as pd

pipe = ChronosPipeline.from_pretrained(
    "amazon/chronos-bolt-base", torch_dtype=torch.bfloat16
)

# historical daily sales
ts = pd.read_csv("sales.csv")["units"].values
context = torch.tensor(ts[-365:])

forecast = pipe.predict(context, prediction_length=30, num_samples=100)
median = forecast.median(dim=1).values  # 30-day median forecast
p90 = forecast.quantile(0.9, dim=1)  # safety stock upper bound
```

### 2. VRP (Vehicle Routing Problem)
```python
from ortools.constraint_solver import pywrapcp, routing_enums_pb2

def solve_vrp(distance_matrix, num_vehicles, depot):
    manager = pywrapcp.RoutingIndexManager(
        len(distance_matrix), num_vehicles, depot
    )
    routing = pywrapcp.RoutingModel(manager)

    def dist_cb(i, j):
        return distance_matrix[manager.IndexToNode(i)][manager.IndexToNode(j)]

    transit_idx = routing.RegisterTransitCallback(dist_cb)
    routing.SetArcCostEvaluatorOfAllVehicles(transit_idx)

    params = pywrapcp.DefaultRoutingSearchParameters()
    params.first_solution_strategy = routing_enums_pb2.FirstSolutionStrategy.PATH_CHEAPEST_ARC
    return routing.SolveWithParameters(params)
```

### 3. (s,S) inventory policy
```python
import numpy as np

def reorder(stock, s, S, demand_forecast, lead_time_days):
    # s = reorder point, S = order-up-to level
    expected_demand_during_lead = demand_forecast.mean() * lead_time_days
    safety = 1.65 * demand_forecast.std() * np.sqrt(lead_time_days)
    s_dynamic = expected_demand_during_lead + safety
    if stock <= s_dynamic:
        return S - stock
    return 0
```

### 4. SBOM generation (CycloneDX)
```bash
# Python project
pip install cyclonedx-bom
cyclonedx-py -o sbom.json --format json

# Node project
npx @cyclonedx/cyclonedx-npm --output-file sbom.json

# Container
syft packages docker:myimage:latest -o cyclonedx-json > sbom.json
```

### 5. Sigstore keyless signing
```bash
# Sign artifact (uses OIDC identity, no long-lived keys)
cosign sign-blob --yes ./release.tar.gz \
  --output-signature release.sig \
  --output-certificate release.crt

# Verify
cosign verify-blob ./release.tar.gz \
  --signature release.sig \
  --certificate release.crt \
  --certificate-identity user@example.com \
  --certificate-oidc-issuer https://github.com/login/oauth
```

### 6. SLSA provenance (GitHub Actions)
```yaml
name: build
on: [push]
permissions:
  id-token: write
  contents: read
jobs:
  build:
    uses: slsa-framework/slsa-github-generator/.github/workflows/generator_generic_slsa3.yml@v2.0.0
    with:
      base64-subjects: ${{ needs.hash.outputs.digests }}
```

### 7. Supplier risk GNN
```python
import torch
from torch_geometric.nn import GraphSAGE

# nodes = suppliers, edges = dependency
model = GraphSAGE(in_channels=16, hidden_channels=64,
                  num_layers=3, out_channels=2)  # risk score

# message passing: tier-1 supplier 의 risk → tier-2 propagation
risk_scores = model(node_features, edge_index)
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| Demand forecast (long horizon) | Chronos / TimesFM |
| Route opt (small, hard) | OR-Tools exact |
| Route opt (large, soft) | RL + heuristic |
| SBOM | CycloneDX (broader) or SPDX |
| Signing | Sigstore (keyless, modern) |

**기본값**: Chronos forecast + OR-Tools VRP + CycloneDX SBOM + Sigstore.

## 🔗 Graph
- 부모: [[Operations-Research]]
- 변형: [[SBOM]] · [[SLSA]] · [[Sigstore]]

## 🤖 LLM 활용
**언제**: Demand pattern 의 explain, anomaly 의 root-cause analysis, SBOM 의 vulnerability summary.
**언제 X**: Real-time route decision (latency), exact optimization (LLM 의 hallucinate cost).

## ❌ 안티패턴
- **Forecast 없는 inventory**: 매 lead time × demand 의 rough-cut estimate → stockout 의 cycle.
- **SBOM 의 build 후 generation**: 매 reproducibility 의 lose. Build 시 generate.
- **Long-lived signing keys**: 매 leak 의 catastrophic. Sigstore keyless 의 use.
- **Dependency 의 pin without lock**: 매 supply-chain attack vector. lockfile + hash check.
- **Tier-1 supplier 의 only monitor**: 매 cascade failure 의 ignore. Multi-tier visibility.

## 🧪 검증 / 중복
- Verified (CSCMP definitions, NIST SSDF SP800-218, SLSA spec v1.0).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — supply chain (logistics + software security) full canonical |