PBX Market Estimation & VoIP Trend Forecasting
dennislee928/pbx_estimation
Previsão quantitativa de declínio de PBX e adoção VoIP por país
Ver no GitHubMCP server que aplica análise de séries temporais e modelos de sobrevivência para prever o declínio do mercado de PBX tradicional e adoção de VoIP em 12+ países. Utiliza modelos logísticos de crescimento (S-Curve) e Cox Proportional Hazards para estimar quando mercados se saturarão e a probabilidade de sucesso de novos produtos em mercados em encolhimento.
Instalação
npx convez-ai dennislee928/pbx_estimation1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Prever saturação/morte do mercado de PBX tradicional por país
- Calcular curvas de sobrevivência de produtos gateway PBX
- Analisar impacto de políticas de desativação PSTN em mercados
- Benchmarking de penetração VoIP/broadband entre regiões
Veredito

Projeto de data science maduro e bem documentado que combina forecasting temporal com análise de sobrevivência. Ideal para profissionais em telecom/infraestrutura que precisam de previsões baseadas em modelos matemáticos rigorosos sobre transição PBX→VoIP.
README
A tradução para o português deste README ainda está na fila.
PBX Market Estimation & VoIP Trend Forecasting
English | 中文
Overview
Quantitative forecasting of PBX market decline and VoIP adoption across 12+ countries using time-series analysis and survival models.
The Problem
Traditional PBX (Private Branch Exchange) systems are being rapidly replaced by VoIP/Cloud-based solutions. This project applies data science to answer two core questions:
- Trend by Country — When will the traditional PBX market die in each country? (Logistic Growth / S-Curve)
- Product Success Rate — What is the survival probability of a new PBX gateway product launched in a shrinking market? (Cox Proportional Hazards Model)
Mathematical Framework
1. S-Curve (Logistic Growth Model) — Trend Prediction
$$P(t) = \frac{K}{1 + e^{-r(t - t_0)}}$$
| Parameter | Meaning |
|---|---|
| $P(t)$ | Market penetration at time $t$ |
| $K$ | Carrying capacity (maximum market size) |
| $r$ | Growth/decay rate |
| $t_0$ | Inflection point (fastest change) |
Implemented via scipy.optimize.curve_fit per country. Outputs: saturation/death year for each market.
2. Cox Proportional Hazards Model — Product Survival
$$h(t|X) = h_0(t) \exp(\sum_{i=1}^n \beta_i X_i)$$
| Parameter | Meaning |
|---|---|
| $h(t|X)$ | Hazard (failure risk) at time $t$ given covariates $X$ |
| $h_0(t)$ | Baseline hazard |
| $X_i$ | Covariates (PSTN phase-out policy, broadband penetration, etc.) |
| $\beta_i$ | Weight of each covariate |
Implemented via lifelines.CoxPHFitter. Outputs: survival curves and hazard ratios per country.
Countries Covered
| Region | Countries |
|---|---|
| Asia-Pacific | Taiwan, Japan, South Korea, China, India |
| Europe | UK, Germany, France, Sweden, Italy |
| Americas | US, Canada, Brazil |
Data Sources
| Source | Data | Method |
|---|---|---|
| World Bank API | Fixed telephone subscriptions, broadband penetration | wbgapi |
| ITU ICT Indicators | IP communication share, infrastructure scores | ITU DataHub API (api.datahub.itu.int) |
| BEREC (European Regulators) | Copper switch-off dates per country | PDF scraping (Table 3) |
| UK House of Commons Library | UK PSTN switch-off timeline | PDF/HTML scrape (CBP-9471) |
| CEPT | European IP migration status | ECC Report 265 PDF scrape |
| NCC (Taiwan) | Local telecom statistics | NCC API + MODA open data (data.gov.tw) |
Project Structure
pbx_estimation/
├── data/
│ ├── raw/ # API downloads cache
│ └── processed/ # Cleaned panel data
├── notebooks/
│ ├── 01_fetch_data.ipynb # Data collection
│ ├── 01b_fetch_supplementary_sources.ipynb # ITU / UK / CEPT / NCC crawlers
│ ├── 02_eda_visualization.ipynb # Exploratory analysis
│ ├── 03_logistic_growth.ipynb # S-Curve fitting
│ └── 04_survival_analysis.ipynb # CoxPH model
├── src/
│ ├── data/fetcher.py
│ ├── data/supplementary_fetcher.py
│ ├── data/preprocessor.py
│ ├── models/logistic_growth.py
│ └── models/survival.py
├── tests/
├── .github/workflows/report.yml # Auto-generate reports
├── config.yaml
├── requirements.txt
└── pyproject.toml
Getting Started
# Clone & install
git clone https://github.com/dennis-lee/pbx_estimation.git
cd pbx_estimation
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt
# Run notebooks
jupyter notebook notebooks/
Report Generation via GitHub Actions
Reports are automatically generated and published via a scheduled GitHub Actions workflow:
- Schedule: Runs every hour at minute 0 UTC
- Trigger: Also supports manual dispatch via GitHub UI
- Steps:
- Fetch latest data from World Bank / ITU / BEREC APIs
- Execute all Jupyter notebooks in order
- Render notebooks to HTML and PDF
- Archive reports as build artifacts (downloadable from Actions tab)
- Optionally deploy to GitHub Pages for a live dashboard
💡 Manual trigger: Go to
Actions→Report Generation→Run workflow→Run now
Cloud RAG Endpoint and Asset Sync
The technology alternatives page uses browser-side keyword filtering for quick narrowing, then calls the Cloudflare Worker in rag_engine/ for cloud RAG prioritization. CI builds a RAG asset manifest from reports/, data/processed/, and frontend/data/, uploads those assets to Cloudflare R2 when configured, and the Worker retrieves report/data evidence from that bucket.
Set these repository secrets for Cloudflare RAG:
CLOUD_RAG_ENDPOINTCLOUDFLARE_API_TOKENCLOUDFLARE_ACCOUNT_IDCLOUDFLARE_R2_BUCKET(current Worker binding expectsauto-rag)- Optional S3 API upload secrets:
CLOUDFLARE_R2_S3_ENDPOINT,CLOUDFLARE_R2_S3_ACCESS_KEY_ID,CLOUDFLARE_R2_S3_SECRET_ACCESS_KEY
Alternatively, the same self-developed RAG engine can run on Hugging Face Spaces as a Docker Space using the files in rag_engine/. Set HF_TOKEN and HF_SPACE_ID to let CI upload the Docker Space. Hugging Face's default CPU Basic Space is currently free and provides 2 vCPU, 16 GB RAM, and 50 GB non-persistent disk. The Space endpoint can also be used as CLOUD_RAG_ENDPOINT.
Cloudflare DNS for GitHub Pages
DNS records cannot point to a URL path. In Cloudflare DNS, do not set a CNAME target such as dennislee928.github.io/pbx_estimation; the /pbx_estimation part is a path, not a hostname.
For a custom subdomain such as mtk-pbx-estimation.example.com, set:
- Type:
CNAME - Name:
mtk-pbx-estimation - Target/content:
dennislee928.github.io
Then set GitHub repository variables:
PAGES_CUSTOM_DOMAIN=mtk-pbx-estimation.example.com- Optional override:
NEXT_PUBLIC_SITE_BASE_PATH=/
When PAGES_CUSTOM_DOMAIN is set and NEXT_PUBLIC_SITE_BASE_PATH is blank, the workflow automatically builds frontend assets for root-domain hosting. Keep NEXT_PUBLIC_SITE_BASE_PATH=/pbx_estimation only when you explicitly want the default https://dennislee928.github.io/pbx_estimation/ URL.
See docs/cloudflare_pages_dns.md for the exact Cloudflare DNS records.
The endpoint returns:
{
"recommendation": "short explanation",
"alternatives": [{ "name": "MQTT (MQTT-SN)", "rank": 1, "reason": "why it fits" }],
"solutions": [{ "name": "Twilio Programmable Voice", "rank": 1, "reason": "why it fits" }],
"documents": [{ "name": "reports/global_research_report_zh.md", "rank": 1, "excerpt": "retrieved report/data evidence" }]
}
NotebookLM
Google NotebookLM does not have an official public file-upload API. CI therefore creates a NotebookLM-ready bundle at rag_engine/dist/notebooklm_sources/ and includes it in the workflow artifact for manual upload. If you run an unofficial bridge such as notebooklm-rest-api or notebooklm-py, set NOTEBOOKLM_UPLOAD_URL and optionally NOTEBOOKLM_API_TOKEN; CI will post the selected sources to that endpoint.
How to Fill .env.example
Copy .env.example to .env for local work, and add the same names as GitHub repository secrets when CI needs them.
| Variable | How to get the value |
|---|---|
NEXT_PUBLIC_CLOUD_RAG_ENDPOINT | Public RAG endpoint used by the frontend build. Use either your Cloudflare Worker URL, for example https://pbxanalyze.pcleegood.workers.dev/, or your Hugging Face Space URL, for example https://<user>-pbx-rag-engine.hf.space/. |
CLOUD_RAG_ENDPOINT | Same endpoint as above, stored as a GitHub secret so .github/workflows/report.yml can pass it into NEXT_PUBLIC_CLOUD_RAG_ENDPOINT during the Pages build. |
NEXT_PUBLIC_SITE_BASE_PATH | Optional GitHub repository variable. Leave blank when PAGES_CUSTOM_DOMAIN is set; the workflow will use /. Use /pbx_estimation only for the default GitHub Pages project URL. |
PAGES_CUSTOM_DOMAIN | GitHub repository variable. Set to the custom domain hostname only, for example mtk-pbx-estimation.example.com; do not include https:// or /pbx_estimation. |
RAG_ASSET_PREFIX | Usually latest. Change only if you want separate R2/Hugging Face asset namespaces such as staging or a dated prefix. |
CLOUDFLARE_API_TOKEN | Cloudflare R2 Account API Token, for example your pbx_application_token. This is used by Wrangler remote upload. It is not the same as an R2 S3 Access Key ID. |
CLOUDFLARE_ACCOUNT_ID | Cloudflare Dashboard → select account → account ID in the dashboard/API Tokens page, or run npx wrangler whoami after login. |
CLOUDFLARE_R2_BUCKET | Create an R2 bucket in Cloudflare Dashboard → R2. Current default: auto-rag. |
CLOUDFLARE_R2_S3_ENDPOINT | Optional S3 API path. If you use only pbx_application_token, leave all S3 fields empty. For S3 API mode, use https://8dfc8c4994bd0925c72ab9e2eff79b48.r2.cloudflarestorage.com/auto-rag; CI strips the trailing /auto-rag when calling the S3 API. |
CLOUDFLARE_R2_S3_ACCESS_KEY_ID | Cloudflare Dashboard → R2 → Manage R2 API Tokens → create an R2 token with object read/write access for the auto-rag bucket. Use the 32-character Access Key ID. Do not use the Cloudflare API token or token id here. |
CLOUDFLARE_R2_S3_SECRET_ACCESS_KEY | Secret Access Key shown once when creating the R2 S3 API token. Store only as a local .env value or GitHub secret. |
HF_TOKEN | Hugging Face → Settings → Access Tokens → create a token with write access to Spaces. Use this as a GitHub secret for CI deployment. |
HF_SPACE_ID | Hugging Face Space repo id in <username-or-org>/<space-name> format, for example dennis-lee/pbx-rag-engine. Create it in the Hugging Face UI or let CI create it with hf repos create ... --type space --space-sdk docker --exist-ok. |
NOTEBOOKLM_UPLOAD_URL | Optional. Only set this if you run an unofficial NotebookLM bridge such as notebooklm-rest-api or notebooklm-py. Use that bridge's upload endpoint URL. Leave empty for manual NotebookLM upload. |
NOTEBOOKLM_API_TOKEN | Optional token required by your unofficial NotebookLM bridge, if any. Leave empty if the bridge does not require bearer auth. |
NOTEBOOKLM_TIMEOUT | HTTP timeout in seconds for the optional NotebookLM bridge upload. Default: 60. |
PORT | Runtime port for the Hugging Face Docker Space/local server. Hugging Face Docker Spaces should use 7860. |
HOST | Runtime host bind address. Use 0.0.0.0 for Docker/Hugging Face Spaces. |
ALLOWED_ORIGINS | CORS allowlist for the RAG engine. Use * for quick testing, or the GitHub Pages origin for production. |
HF_RAG_ASSET_ROOT | Local asset directory inside the Hugging Face Docker container. Keep /app/dist/hf_assets unless you change the Dockerfile. |
USE_WORKERS_AI | Cloudflare Worker setting. Use true to call Workers AI for the final summary, or false to use deterministic retrieval/ranking only. Hugging Face Docker Space uses deterministic mode. |
Dependencies
pandas, numpy, scipy, matplotlib, seaborn, wbgapi, lifelines, scikit-learn, pyyaml, requests, jupyter, nbconvert, papermill
概述(中文)
多國 PBX 市場衰退與 VoIP 普及趨勢的量化預測專案,採用時間序列分析與存活分析模型。
問題
傳統 PBX 交換機正被 VoIP/雲端方案快速取代。本專案透過資料科學回答兩個核心問題:
- 各國趨勢預測 — 各國傳統 PBX 市場何時消亡?(Logistic S 曲線模型)
- 產品成功率評估 — 在萎縮市場推出新 PBX 網關產品的存活機率為何?(Cox 比例風險模型)
數學模型
1. S 曲線(羅吉斯成長模型,又稱邏輯性成長模型)— 趨勢預測
$$P(t) = \frac{K}{1 + e^{-r(t - t_0)}}$$
使用 scipy.optimize.curve_fit 對每個國家獨立擬合,預測市場飽和點與消亡時間點。
2. Cox 比例風險模型 — 產品存活分析
$$h(t|X) = h_0(t) \exp(\sum_{i=1}^n \beta_i X_i)$$
使用 lifelines.CoxPHFitter,共變量包含 PSTN 退場政策(BEREC 資料)、寬頻普及率等。
資料來源
- World Bank API — 各國固定電話訂閱數歷年資料
- ITU DataHub API — ICT 指標(固網、寬頻、網際網路使用者)
- BEREC(歐洲監管機構聯盟)— 各國銅纜退場時程表(PDF 解析)
- UK Parliament — 英國 PSTN 關閉時間軸(Commons Library CBP-9471)
- CEPT — 歐洲 PSTN/ISDN 轉 IP 遷移報告(ECC Report 265)
- NCC(台灣) — 行動通信用戶統計、市話號碼核配(NCC API / 政府資料開放平臺)
使用方式
git clone https://github.com/dennis-lee/pbx_estimation.git
cd pbx_estimation
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt
jupyter notebook notebooks/
GitHub Actions 自動報告產出
透過 GitHub Actions 定期自動執行:
- 排程:每小時 UTC 整點自動執行
- 手動觸發:GitHub UI → Actions → Report Generation → Run workflow
- 流程:抓取最新資料 → 依序執行所有 Notebook → 輸出 HTML/PDF 報告 → 產出可下載的 Artifact
雲端 RAG 端點與資產同步
技術替代方案頁面只在瀏覽器做快速關鍵字篩選,優先排序由 rag_engine/ 的 Cloudflare Worker 回傳。CI 會將 reports/、data/processed/、frontend/data/ 建成 RAG asset manifest,並在設定 Cloudflare secrets 後上傳到 Cloudflare R2 bucket,Worker 會從 bucket 讀取報告/資料證據。
需要的 repository secrets:CLOUD_RAG_ENDPOINT、CLOUDFLARE_API_TOKEN、CLOUDFLARE_ACCOUNT_ID、CLOUDFLARE_R2_BUCKET。
也可以用 rag_engine/ 的 Dockerfile 將同一套自研 RAG engine 部署到 Hugging Face Spaces。設定 HF_TOKEN 與 HF_SPACE_ID 後,CI 會上傳 Docker Space;Hugging Face CPU Basic Space 目前為免費方案,並提供 2 vCPU、16 GB RAM、50 GB 非持久磁碟。Space URL 也可作為 CLOUD_RAG_ENDPOINT。
NotebookLM
Google NotebookLM 沒有官方公開檔案上傳 API。CI 會產生 rag_engine/dist/notebooklm_sources/,並放進 workflow artifact 供手動上傳。若你自行架設 notebooklm-rest-api 或 notebooklm-py 這類非官方橋接服務,可設定 NOTEBOOKLM_UPLOAD_URL 與選用的 NOTEBOOKLM_API_TOKEN,CI 會將來源檔 POST 到該端點。
如何填寫 .env.example
本機開發可將 .env.example 複製為 .env;CI 需要使用的值,請用同名 GitHub repository secrets 設定。
| 變數 | 取得方式 |
|---|---|
NEXT_PUBLIC_CLOUD_RAG_ENDPOINT / CLOUD_RAG_ENDPOINT | 使用 Cloudflare Worker URL 或 Hugging Face Space URL。前者給前端建置用,後者是 CI secret 名稱。 |
RAG_ASSET_PREFIX | 通常維持 latest。只有要分 staging/date namespace 時才需要修改。 |
CLOUDFLARE_API_TOKEN | Cloudflare R2 Account API Token,例如你的 pbx_application_token。這是給 Wrangler remote upload 用,不是 R2 S3 Access Key ID。 |
CLOUDFLARE_ACCOUNT_ID | Cloudflare 帳號頁面取得,或登入後執行 npx wrangler whoami。 |
CLOUDFLARE_R2_BUCKET | Cloudflare R2 建立 bucket;目前預設 auto-rag。 |
CLOUDFLARE_R2_S3_ENDPOINT | 選填的 S3 API 路徑。如果只使用 pbx_application_token,請讓所有 S3 欄位保持空白。若使用 S3 API 模式,你的 bucket 可填 https://8dfc8c4994bd0925c72ab9e2eff79b48.r2.cloudflarestorage.com/auto-rag。 |
CLOUDFLARE_R2_S3_ACCESS_KEY_ID / CLOUDFLARE_R2_S3_SECRET_ACCESS_KEY | Cloudflare Dashboard → R2 → Manage R2 API Tokens 建立限定 auto-rag bucket 的讀寫 token;Access Key ID 是 32 字元,不是 Cloudflare API token 或 token id。secret key 只會顯示一次,放 GitHub secret。 |
HF_TOKEN | Hugging Face Settings → Access Tokens 建立具 Spaces 寫入權限的 token。 |
HF_SPACE_ID | Hugging Face Space repo id,格式 <username-or-org>/<space-name>。 |
NOTEBOOKLM_UPLOAD_URL / NOTEBOOKLM_API_TOKEN | 只有使用非官方 NotebookLM bridge 時才填;手動上傳 NotebookLM 時保持空白。 |
NOTEBOOKLM_TIMEOUT | NotebookLM bridge 上傳 timeout 秒數,預設 60。 |
PORT / HOST | Docker/Hugging Face Spaces 使用 7860 與 0.0.0.0。 |
ALLOWED_ORIGINS | RAG engine CORS 設定;測試可用 *,正式環境建議填 GitHub Pages origin。 |
HF_RAG_ASSET_ROOT | Hugging Face Docker container 內的資產路徑,除非改 Dockerfile,維持 /app/dist/hf_assets。 |
USE_WORKERS_AI | Cloudflare Worker 是否使用 Workers AI 產生摘要;Hugging Face Docker Space 使用 deterministic ranking。 |
License
MIT
CI/CD Crawler Summary
Last generated: 2026-06-09
- 155 解決方案
- 42 國家/地區
- 131 替代技術
- 126 供應商