惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

月光博客
月光博客
云风的 BLOG
云风的 BLOG
小众软件
小众软件
雷峰网
雷峰网
博客园 - 【当耐特】
V
V2EX
WordPress大学
WordPress大学
IT之家
IT之家
Last Week in AI
Last Week in AI
罗磊的独立博客
奇客Solidot–传递最新科技情报
奇客Solidot–传递最新科技情报
Apple Machine Learning Research
Apple Machine Learning Research
钛媒体:引领未来商业与生活新知
钛媒体:引领未来商业与生活新知
V
Visual Studio Blog
freeCodeCamp Programming Tutorials: Python, JavaScript, Git & More
有赞技术团队
有赞技术团队
The Cloudflare Blog
Jina AI
Jina AI
博客园 - 司徒正美
阮一峰的网络日志
阮一峰的网络日志
博客园 - 聂微东
大猫的无限游戏
大猫的无限游戏
博客园 - 三生石上(FineUI控件)
让小产品的独立变现更简单 - ezindie.com
让小产品的独立变现更简单 - ezindie.com

DEV Community

Authentication Security Deep Dive: From Brute Force to Salted Hashing (With Java Examples) Why AI Systems Don’t Fail — They Drift Spilling beans for how i learn for exam😁"Reinforcement Learning Cheat Sheet" I Replaced Chrome with Safari for AI Browser Automation. Here's What Broke (and What Finally Worked) How Python Borrows Other People's Work The $40 Architecture: Processing 1 Billion API Requests with 99.99% Uptime Vibe Coding: A Workflow Guide (From Zero to SaaS) Most webhook security guides protect the wrong side. The scary part is delivery. Headless CMS for TanStack Start: Build a Blog with Cosmic EU Age Verification App "Hacked in 2 Minutes" — What Actually Happened Comfy Cloud’s delete function does not actually remove files Running AI Models on GPU Cloud Servers: A Beginner Guide Event-driven media intelligence with AWS Step Functions and Bedrock I scored 500 AI prompts across 8 quality dimensions — here's what broke How to Call Google Gemini API from Next.js (Free Tier, No Backend Needed) The Portal Protocol: Reclaiming Human Connection in the Age of AI How to Fix Your Team's Scattered Knowledge Problem With a Self-Hosted Forum Intro to tc Cloud Functors: A Graph-First Mental Model for the Modern Cloud Designing Multi-Tenant Backends With Both Ownership and Team Access I Built a Neumorphic CSS Library with 77+ Components — Here's What I Learned PostgreSQL Performance Optimization: Why Connection Pooling Is Critical at Scale Cómo construí un SaaS multi-rubro para gestionar expensas en Argentina con FastAPI + Vue 3 🚀 I Built an Ethical Hacking Scanner Tool – Open Source Project I Replaced /usage and /context in Claude Code With a Single Statusline A Pythonic Way to Handle Emails (IMAP/SMTP) with Auto-Discovery and AI-Ready Design I Collected 8.9 Million Polymarket Price Points — Here's What I Found About How Markets Really Move EcoTrack AI — Carbon Footprint Tracker & Dashboard Everyone's Using AI. No One Agrees How. 5 self-hosted ebook managers worth trying in 2026 Building Your First AI Agent with LangChain: From Chatbot to Autonomous Assistant
Build Rails, Not Trains: A Framework for AI Infrastructur...
Gabriel Mahia · 2026-06-21 · via DEV Community

Build Rails, Not Trains: A Framework for AI Infrastructure in the Global South

There's a question I ask before building anything:

"What is missing?"

Not: "How do I compete with what already exists?"

The answer to the second question leads you toward incremental improvement.
The answer to the first question leads you toward infrastructure.


The Rail vs. The Train

TCP/IP didn't control the internet. It enabled it.
Railroads didn't control freight. They enabled an economy.
M-Pesa didn't control mobile money. It enabled 50 million transactions a month.

Rails created ecosystems. The trains came later — built by thousands of people
who didn't need to understand the rails to use them.

For the past several months I've been building rails.

Not apps. Not chatbots. Not wrappers.

31 MCP servers that give AI agents structured, authenticated, locally-processed access to:
the M-PESA API, Kenya's 47-county government layer, NDMA drought classifications across all counties,
land title systems, health infrastructure, education records, SACCO finance, matatu routes, and more.

Any developer — anywhere — can now do this:

pip install mpesa-mcp county-mcp wapimaji-mcp kilimo-mcp

And give an AI agent the ability to trigger a mobile payment, query county budget data,
check drought severity by county, or assess agricultural market prices.
In under 120 seconds. On any LLM. No API key required beyond Daraja credentials.

That's the rail. You build the train.


Diagnose Structural Absences

The standard move in tech is competitive analysis.
You look at what exists, identify what's better or cheaper, and build that.

That's fine for markets with existing demand.

For East Africa's institutional AI layer, the problem wasn't competition. It was absence.

M-Pesa wasn't missing from the MCP ecosystem because engineers hadn't tried.
It was missing because the engineer who understood how it works — its callbacks, its STK Push flow,
its B2C timing, its idempotency requirements — wasn't building there.

Every MCP server in this stack was born from the same question:

"What does an AI agent need to be genuinely useful in this context — and what doesn't exist yet?"

The answers:

  • A way to trigger payments (mpesa-mcp)
  • A way to query drought severity before a farming decision (wapimaji-mcp)
  • A way to check what documents a government process requires (fomu-mcp)
  • A way to run inference with no internet and no API key (offline-mcp)

None of these existed. Now they do.


Build at Your Intersections

The technical knowledge to build MCP servers isn't rare.
The knowledge of how M-Pesa works from the inside isn't rare either.
The knowledge of Kenya's 47-county government structure is widely shared.

What's rare is the intersection.

Kenyan + diaspora + AI infrastructure fluency.

That intersection doesn't have a lot of occupants.
It doesn't need many. It just needs someone to start building.

Every person has an intersection like this.
The place where what you know uniquely, what the world needs, and what the moment makes possible all overlap.

That's not a motivational statement. It's a resource allocation principle.
Build where others can't easily replicate you.


The Timing Layer

Not all periods are equal.

In 2003, you couldn't have built a mobile payment layer in Kenya — the infrastructure didn't exist.
In 2013, you could, but you'd have needed to build the whole stack.
In 2023, M-Pesa had a mature API, Africa's Talking existed, and the missing piece was AI coordination.

The AI infrastructure window is open right now.
The MCP protocol exists. The LLM APIs exist. The institutional data exists.
What's missing is the coordination layer for contexts that aren't Silicon Valley.

Windows close.

I'm not claiming certainty about when. But the pattern from previous infrastructure cycles
(telecom, mobile, cloud) is clear: the window for foundational protocol-level work is short.
Execution matters more than perfection during that window.


Stewardship, Not Ownership

There's a version of this work that looks like ownership.
Build the M-Pesa MCP server, gate access behind a subscription, own the coordination layer.

That's a train, not a rail.

The entire portfolio is MIT licensed.
The SII Stack has a sovereign inference tier
(Ollama, local, free) so communities can run AI without any data leaving their device.
The offline-mcp server runs Llama 3.2
on a Raspberry Pi with no internet and no API key.

Why? Because the deepest principle in this work isn't efficiency or scale.

It's stewardship.

Across the Global South, communities are increasingly pressured to hand over health data,
land records, and civic information as conditions for receiving services or funding.
The architecture of this infrastructure was designed so communities can deliver
AI-powered services without surrendering that data.

You don't own infrastructure. You steward it.


The Decay Function

Not everything compounds.

Skills decay. Platforms decay. Relevance decays.
Character decays slowly. Judgment decays slowly. Intersectional expertise decays slowly.

The 31 MCP servers aren't apps. Apps decay fast — a new framework, a changed API,
a shifting platform, and months of work can become worthless.

Infrastructure decays slowly. TCP/IP is 50 years old. M-Pesa is 17 years old.
The MCP protocol is new, but the pattern is established: protocols outlive implementations.

When choosing what to build, ask: what's the decay rate?


What the Stack Looks Like Now

31 MCP servers  →  pip install {server-name}
SII Stack       →  n8n + LiteLLM (tri-polar) + Ollama + Postgres
                   Western / Eastern / Sovereign routing
                   72-hour offline test: must work without internet
5 HF datasets   →  246 total downloads
15 Dev.to articles → 257 total views

The rail exists. The trains are next.

If you're building AI tools for contexts outside the default infrastructure assumptions —
Global South, rural, offline-first, sovereignty-constrained, institutional rather than consumer —
these servers are available to use, fork, extend, or build on.

MIT licensed. No conditions.


Full portfolio
GitHub
PyPI
Glama