When Hammad Malik built a simple Urdu WhatsApp bot and gave it to his driver to try out, he didn’t expect 800 people to be using it within two days. That early experiment showed him that millions of Pakistanis are excluded from digital services because they cannot interact with technology in their own languages.

Uplift AI was born from this insight. Founded in March 2024 by Malik and fellow former Apple and Amazon engineer Zaid Qureshi, the startup is building voice-based technology that lets people speak to devices in their native languages. The company has announced that it has raised $3.5 million in funding from Y Combinator, Indus Valley Capital and others to expand its work.

The problem Uplift AI is trying to address is structural. According to World Bank data, around 40 per cent of Pakistanis aged 15 and above cannot read or write a short, simple sentence about everyday life, which effectively excludes them from text-based systems.

Still in its early stages, Uplift AI’s flagship model, Orator, currently offers services in Urdu, Sindhi and Balochi. The company follows a business-to-business model and has two early customers: Khan Academy, which has used the technology to release 2,500 educational videos in Urdu, and the agricultural chemicals company Syngenta, which is building technology to assist farmers in local languages using Uplift AI’s models.

Uplift AI began with Balochi and Sindhi deliberately. The aim was to prove the technology using the hardest languages with the least available data. Balochi, in particular, posed a major challenge: less than 5pc of Pakistan’s population speaks it, and even fewer can write it, leaving almost no usable data online. As a result, the company had to generate entirely new, local data sources from scratch.

The startup’s initial seed funding came from friends and family. “My friends gave me $5,000 each, and I raised about $125,000 back in November,” says Malik. After moving back to Pakistan from the US, he began collecting voice data in local languages to train the models. From listening to workers speak in textile mills to conversations in farmlands, Uplift AI built its Urdu, Sindhi and Balochi datasets. Punjabi, Pashto and Saraiki are scheduled to roll out later this year, with Punjabi expected first, given it is spoken by the country’s largest farming cohort and forms a key part of the company’s work with Syngenta.

The gap between language and technology is especially visible in agriculture. Syngenta employs nearly 1,000 people whose sole job is to respond to farmers’ voice-based queries, says Malik. The issue is not the size of the workforce, but inconsistency. As with most customer service operations, the quality of responses varies widely, often leading to incorrect use of agricultural products and, ultimately, lower yields. Voice AI, he argues, can play a transformative role by providing clear guidance in a language farmers understand, thus improving crop productivity.

Many mothers have heaved a sigh of relief, now that ChatGPT can help with homework in Urdu as well, complete with voice input. Does that make large language models a competition for Uplift AI?

Not really, says Malik. Emerging markets are not a priority for global technology companies. Their language capabilities are often superficial, limited to what can be scraped from the internet. Uplift AI, by contrast, uses OpenAI and other providers only for the large-language-model layer, the cognitive understanding of what a user is asking, while building its own voice infrastructure for local languages.

Most global AI models rely on millions of hours of internet data, which is either manually labelled or generated by automated systems. But for many Pakistani languages, there is no baseline to work from at all.

Of the $3.5m raised, more than $1m will be spent in Pakistan to generate original voice data across regional languages. This includes collecting recordings from people across diverse environments and hiring thousands to label them, a process Malik believes is essential if voice AI is to work meaningfully for Pakistan.

With devices such as VR glasses, wearables, and robots becoming mainstream, he argues, voice will become the primary way humans interact with machines. And when that technology becomes affordable enough to reach the rest of the world, language will matter. “People will prefer to talk to their technology in the language they know. If someone speaks Pashto, they would want to speak in Pashto to their robots.”

Uplift AI’s vision is to provide this layer to global hardware and platform companies, from robotics firms to VR manufacturers such as Meta or Apple. “We want to be the best in the world for languages no one else wants to build for,” he says.

However, challenges are significant. Voice AI remains expensive even in developed markets, to the point where some US companies limit its use. Deploying it in developing countries, where purchasing power is a fraction of that in the West, makes the equation even harder.

After the spate of startups crashing and burning in recent years, the question of sustainable profitability for expensive technology that has large potential but only in theory inevitably arises. Malik acknowledges the concern but argues that startup failure is not unique to Pakistan. “In Silicon Valley, Y Combinator funds over a hundred startups every few months, and only around 6pc become massive successes,” he says. “You only hear about the successful ones.”

For Malik, public services and commerce remain among the strongest use cases for voice AI, precisely because these systems are heavily text-based and currently exclude large segments of the population.

Currently, the startup is focusing on businesses willing to pay rather than offering a free service to end consumers, ensuring variable costs are covered. The funding is for the R&D stage of which costs can be recouped if Uplift AI can scale it. Startups, he adds, are high-risk by definition wherever you go. The challenge is to ensure that capital is not just spent, but converted into sustainable value before the runway runs out.

Source link