Let’s get the expectations out of the way first. This is not a product launch, a corporate press release, or a venture-backed pitch. It is simply a builder’s log from a weekend spent tinkering.
I wanted to see if I could build a completely local, privacy-safe automated AI agent to parse, supervise, and verify Kenyan tax documents—specifically KRA PIN Certificates—offline on my own machine . I didn’t have a funded lab or a massive database. I had a laptop, open-source models running via Ollama, and a tiny test registry of exactly twenty-four documents (n=24).
It is a pet project, a sandbox experiment, and a way to get my hands dirty with code after a long, bruising year. Here is how the pipeline works, why I built it, and why the quiet, unglamorous act of building is sometimes its own reward;
The Concrete on the Site
Meja Mwangi opens his novel Going Down River Road on a Nairobi construction site. Ben and Ocholla, anonymous workers, pouring concrete for a building that will one day have its name on a brass plate outside, while theirs appear nowhere. The city needs the building. The building needs the concrete. They show up every morning regardless — because the work is the work, and because no one else is going to pour it for them.
This is a builder’s log, written from the middle of a weekend build.
The year that preceded this project had done everything a year can do to remove a person from the game. Court judgments. The particular silence of people who called themselves allies choosing comfort over conviction. I had every reason to go quiet. To regroup in private. To wait for the noise to pass before showing up again.
Instead, I opened a laptop and started tinkering.
Not from a funded lab. Not from a corporate research team with quarterly OKRs. From a desk, with open-source models, NO GPU/s, and twelve years of knowing exactly what it costs to stay compliant in Kenya when you are the founder, the accountant, and the compliance officer simultaneously. Like Ben and Ocholla, I was not waiting for the city to invite me to build. I was already on site.
This essay is about one item. There are many more being built simultaneously on the same foundation. But this one I could not walk past.
The Problem That Wouldn’t Let Me Rest
There is a particular kind of exhaustion that only finds you at 11pm on the eve of a VAT filing deadline.
You know the feeling. The iTax portal is loading—and has been loading—for the past four minutes. The spreadsheet on your second screen holds months of receipts that a real accountant would have summarized in a week, but you are not a real accountant. You are a founder, which means you are also the tax researcher, the compliance officer, and the person who Googles “withholding tax rate on professional fees Kenya” at midnight because the official portal is behaving erratically.
The cost of staying compliant in Kenya is not low. And it is highest precisely for those who can least afford to carry it.
In Bessie Head’s short story Looking for a Rain God, a Botswana family endures a devastating seven-year drought. The rain gauges stay empty, the formal systems offer no relief, and eventually, desperation drives them to reach back into forbidden, ancient knowledge just to make the ground yield. The story ends in tragedy, but the underlying impulse is something every small business owner sitting in front of a spinning iTax loading screen understands: when the official system fails you long enough, you stop waiting for it to save you. You start looking for your own rain god.
The Machine Kenya Is Buying — And the One I Am Building
In 2025, the Kenya Revenue Authority issued Tender No. KRA/HQS/NCB-036/2025-2026. They called it Ushuru GPT- a cloud-hosted, large language model system designed for taxpayer self-service, document ingestion, and automated compliance handling
It is an ambitious, necessary vision. But it brought a critical question to my mind: who builds it?
If we rely entirely on imported enterprise systems, the vendor who wins the contract will certainly localize it, but they won’t understand the ground-level texture. They won’t know that a KRA PIN Certificate issued by the Mombasa branch might spell a taxpayer’s name slightly differently from the Nairobi registry. They won’t know that the moment a Kenyan SME owner interacts with a chatbot about their taxes, the first thing they need is not a fancy feature—it is trust.
Starting from the Bottom: n=24
Silicon Valley purists who are obsessed with “big data” will look at n=24 and scoff. They are thinking about scale before they even understand the problem. In bootstrapping, you do not wait for a million data points to begin; you start with whatever data you can actually get your hands on
My experimental tax intelligence sandbox currently has exactly twenty-four ground truth records – twenty-four verified KRA certificates, parsed, supervised, and matched against KRA database.
These twenty-four records gave me a playground—an evaluation harness to measure latency, hallucination, and OCR degradation.
It reminds me of Emmanuel Dongala’s short story, The Man . A nameless villager walks into a heavily fortified capital city and accomplishes what Western security experts swore was impossible. He did not bring an army or superior firepower. He succeeded because he had hyper-localized resolve and local intelligence. He knew the terrain.
Specialization Without Isolation: Gemma and Qwen
To process this small dataset, I set up a simple pipeline run on a principle familiar to anyone who has ever managed a Kenyan team well: specialization without isolatio
I am using two local, open-source models running on-device via Ollama. Gemma 2 is the extractor; Qwen 2.5 is the supervisor
The final verification doesn’t rely on AI at all. It queries KRA’s GavaConnect API—an integration we had already built and tested for our onboarding platform, Lattice Finance.
The Sandbox and the Scratchpad
Let’s be entirely clear—this is a toy. It is a pet project, a Proof of Concept (POC). It might eventually find a practical, internal use case to help run compliance inside Pawa IT. Or, it might just as easily end up in our digital garbage tin of abandoned code.
But in the life of a builder, this is never a waste of time or resources. Every line of code written is an invaluable lesson learned, a skill sharpened, and another tool under my belt. Even if this POC goes straight to the bin tomorrow, the engineering muscle remains.
We are not building toward instant perfection; we are just enjoying the quiet grit of the build. Like Ben and Ocholla down on the Nairobi construction sites, we just keep pouring the concrete.
The machine is not waiting for permission. And neither am I.
Oscar Limoke is the co-founder of Pawa IT Solutions and Konversations. He writes about technology, leadership, and the unglamorous reality of building in Africa’s Silicon Savannah.