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