[KongchangAI]
· 2 min read· 1,435 words

Kenyan Startup Fastagger: Running AI on Offline, Low-End Phones

Kenyan Startup Fastagger: Running AI on Offline, Low-End Phones

Kenyan startup Fastagger builds offline AI for low-end phones, bringing business intelligence to small African businesses.

While the global AI race chases larger models and more compute, Kenyan startup Fastagger goes the opposite way, developing AI that runs on offline, low-power phones. Its product Auni is integrated into M-Pesa, Kenya's largest mobile payment service, and has attracted about 6,000 small businesses. Economist Daniel Björkegren warns that a clear divide between small models for emerging economies and large models for wealthy regions could lead to AI capability divergence.

While the global AI race focuses on larger parameter counts and more powerful compute, a startup from Kenya has taken the completely opposite path: stuffing AI into low-end phones so it can work even without a network connection. Fastagger's story shows that the key to AI for all may not lie in how powerful the model is, but in whether it is willing to be redesigned for the most ordinary users.

Kenyan startup Fastagger rethinks AI for small businesses

The overlooked Africa, and a question asked late

In 2019, Mutembei Kariuki attended a conference on the future of AI at the University of Tokyo. Speaker after speaker talked about how the technology would change the world, but Africa was not mentioned once throughout the entire event. "Nobody mentioned Africa," he recalled.

At the time, OpenAI and Anthropic had not yet released chatbots to the public, but Kariuki had already clearly seen the practical obstacles to AI adoption in Africa: networks are slow and uneven, the compute needed to train models is scarce, and high-quality structured training data is almost nonexistent. More critically, most people access the internet through old feature phones with limited memory, not computers or high-end smartphones.

After returning home, Kariuki co-founded Fastagger with Jude Mwenda and Stephanie Njerenga, with a clear goal—to build AI models that can run offline on low-power phones.

Auni: Reading business insights from the mobile wallet

Fastagger's first major product, Auni, is a business intelligence assistant. It is integrated into the business app of M-Pesa, Kenya's largest mobile payment service, giving it crucial distribution support. According to IEEE Spectrum, about 6,000 businesses had signed up to use Auni after launch, including barbershops, restaurants, and small manufacturers.

What these small businesses have in common is the lack of any proper data infrastructure. Peter Chege, the owner of a barbershop in Nairobi, used to write all his sales records on paper. "We couldn't turn that information into charts." After connecting to Auni, he could see which day and which time slot was busiest, and identify regular customers; if a regular stopped showing up, he could send a personalized message to bring them back.

For Chege, the most direct value is time management. A simple chart showed that mornings were the slowest period for business, so he scheduled errands for that time, or delayed adding barbers until foot traffic picked up at 1 p.m.

M-Pesa is Kenya's largest mobile payment service, launched by Safaricom in 2007. Users can transfer money, pay, and save through feature phones without needing a bank account. Its penetration in Kenya is extremely high—more than 80% of adults use it—so it has accumulated massive amounts of transaction data. For small businesses, M-Pesa statements and transaction records effectively constitute their only "financial system," which is why Fastagger chose it as the data entry point.

Designing for real-world conditions, not forcing a product in

Fastagger's approach is fundamentally different from "slightly adapting a product from a mature market." Before developing Auni, the team first solved the question of where the training data would come from. They set their sights on mobile payment apps—more than 80% of adults in Kenya use mobile wallets, which is itself a data goldmine.

The first step was to adapt an open-source OCR model so it could read financial statements generated by mobile payment apps. Off-the-shelf models performed poorly on these low-quality PDFs: unfamiliar formats, plus stamps and errors. Considering that target users have unstable networks, the team moved key steps such as table organization to run on-device.

Next was understanding users' rhythm of life. The team spent whole stretches of time observing how small business owners work and interact with customers. One obstacle quickly emerged: shop owners stand all day cutting hair and simply have neither the time nor the energy to think about anything beyond day-to-day operations. Kariuki said, "You have to think about that dynamic of the user journey, how they live, what happens in the background."

In terms of feature design, Auni layered its capabilities: reading sales statements, analyzing data, and visualizing results can be done offline on low-cost phones without initiating cloud queries; more complex analysis and Q&A around the data require a network connection. This type of functionality blends AI with traditional software—standard OCR extracts text, and a small neural network determines document structure.

OCR (optical character recognition) is a technology that converts text in images into editable text. Traditional OCR models are usually trained on scanned documents or clear printed text, but PDFs generated by mobile payment apps often have irregular formats, low resolution, and interference such as stamps and handwritten annotations, which reduces the recognition accuracy of off-the-shelf models. Fastagger adapted an open-source OCR model to these low-quality documents and avoided unstable networks by running key steps on-device.

Multi-source perspectives: innovation forced by constraints, with divergence risks hidden within

Beyond Auni, Fastagger has also developed a workforce management tool for Kenya's security industry. Security guards are stationed in remote locations without stable networks and also use low-end phones. Fastagger's AI agent communicates with them via SMS, tracks patrol routes, detects anomalies, and generates actionable alerts for operations teams.

Fastagger is not alone. According to IEEE Spectrum, there are applications in Chile for rural schools without internet, a medical triage chatbot in India serving rural patients, and African language translation models that can run on ordinary people's existing phones. These innovations growing out of constraints are advancing the democratization of AI.

Daniel Björkegren, an economist at Columbia University who studies AI's impact on developing economies, believes that sub-Saharan Africa's reality of being dominated by basic phones may give rise to different forms of AI use. "The history of technology should tell us that this won't be people running Claude Code or Codex on laptops," he said. "You'll see those people, but for the majority of the population, we'll have to figure out different ways of using it."

But he also raised a judgment worth heeding: if the boundary between small models designed for emerging economies and the ever-larger models built for wealthy regions becomes too clear, it could lead to "divergence." "We need to provide a rich menu of options, make sure people have access, and then see what actually works."

This divide deserves serious attention. On the optimistic side, engineering innovation forced by constraints can indeed bring AI to more people; on the pessimistic side, if small models are positioned as "good enough" downgraded solutions, the digital divide could become entrenched in new forms.

AI divergence refers to different regions developing vastly different forms of AI applications and capability levels due to differences in infrastructure, data, compute, and other conditions. Wealthy regions tend to pursue general-purpose large models with more parameters and stronger performance, while emerging economies may rely on small, specialized models designed for low-resource environments. If the capability gap between the two continues to widen and there is a lack of interoperability and migration paths, the digital divide could become entrenched in new forms—users of small models could be locked into a "good enough" downgraded experience and unable to enjoy the advanced capabilities of frontier AI.

A three-person company, and a bigger question

Fastagger has received support from accelerator programs run by Google, Microsoft, and NVIDIA, but still has only three employees—the three co-founders. They are adding agent features to Auni that can automatically execute owners' marketing actions.

The value of this small company is not in its scale, but in the example it provides: when an AI product is designed starting from the most constrained user conditions, it may instead reach broader real-world needs. For global AI developers, this may be a direction more worth thinking about than the parameter race.

One-line Summary

Kenyan startup Fastagger builds offline AI for low-end phones, bringing business intelligence to small African businesses.

Paragraph Summary

While the global AI race chases larger models and more compute, Kenyan startup Fastagger goes the opposite way, developing AI that can run on offline, low-power phones. Its product Auni is integrated into M-Pesa, Kenya's largest mobile payment service. By adapting an open-source OCR model to read transaction statements and running key steps on-device, it has attracted about 6,000 small businesses and helps barbershops, restaurants, and others analyze foot traffic and manage time. Fastagger has also launched an SMS-based AI agent for the security industry. Economist Daniel Björkegren warns that if the boundary between small models for emerging economies and large models for wealthy regions is too clear, it could lead to AI capability divergence, requiring a rich menu of options rather than simple downgraded solutions.

#Fastagger#Auni#M-Pesa#offline AI#low-end phones#Kenya AI#AI divergence#OCR
Share:

Related articles