Applied AI in Kenya
I build AI systems that do one job properly — sales agents that qualify and close in WhatsApp, document processors that read what nobody has time to read, and retrieval systems that answer from your own data rather than guessing.
The AI projects that pay for themselves are dull and specific. One bottleneck, one workflow, measured properly. The ambitious ones tend to stay in a document.
What that looks like in practice: an agent that handles an enquiry end to end in WhatsApp — understands the request, quotes, takes payment, and escalates to a human when it should. A document pipeline that extracts structured data from PDFs a person would otherwise retype. A retrieval system grounded in your own content, so answers come from your documents instead of the model's imagination.
I work across OpenAI, Claude, Gemini and open models via OpenRouter, picking per task rather than per preference, and build with tool calling, RAG and MCP where they earn their place. The decision that matters is usually not which model — it is what the system does when the model is wrong.
A narrow system that ships beats a brilliant one that stays in a document. Every AI build starts by naming the manual work it is meant to remove, so there is something to measure afterwards.
What this fixes
- Repetitive knowledge work
- Slow decision-making
- AI ideas that never ship
Work that proves it
- WhatsApp AI Sales Agent with M-Pesa CheckoutAn AI sales agent that runs entirely inside WhatsApp — search a real catalogue, get a grounded recommendation, and pay by M-Pesa — with the owner SMS’d the moment a sale lands.
- WhatsApp Bundle Sale AutomationA WhatsApp bot that sells data bundles and confirms M-Pesa payment end to end.
- Order-to-Dispatch Bot for Boutiques & ShopsA WhatsApp bot that takes the order, takes the money, updates stock and calls the rider.
Tools and platforms
- OpenAI
- Claude
- Gemini
- DeepSeek
- OpenRouter
- LangChain
- MCP
- Prompt Engineering
- AI Agents
- RAG
- Function Calling
- APIs
OpenAI
Claude
Gemini
DeepSeek
OpenRouter
LangChain
MCP
Prompt Engineering
AI Agents
RAG
Function Calling
APIsCommon questions
- What can AI realistically do for a small business?
- Remove specific bottlenecks: answering enquiries the moment they arrive, reading and extracting data from documents, drafting and following up, and qualifying leads before a human spends time on them. It does not replace judgement or relationships, and any proposal that claims it does is worth ignoring.
- Can an AI agent take payments?
- Yes. The WhatsApp AI sales agent in the case studies handles the conversation, triggers an M-Pesa STK push at the right moment, confirms the payment, and hands off to a human on anything it should not decide alone.
- Which AI model do you use?
- Whichever fits the task — OpenAI, Claude, Gemini, DeepSeek, or an open model through OpenRouter. Models are swappable; the system around them is the actual work. Builds are structured so changing the model later is a configuration change, not a rewrite.
- Is my business data used to train the model?
- No. Systems are built against API endpoints with training disabled, and where data sensitivity demands it, the retrieval layer and storage stay under your control.