Former Goldman and Meta Founders Launch Voice AI for Underserved Markets

The voice AI sector is experiencing an unprecedented surge in demand for customer support and services. However, developing products that effectively imitate human interaction and offer quick responses is quite challenging across various markets, especially in Africa and the Middle East—regions often overlooked by major companies initially.

AethexAI, a startup launched last year to fill this gap, has successfully secured $3 million in pre-seed funding, primarily from 4DX Ventures, along with contributions from Enza Capital, Dorm Room Fund, Mojo Ventures, and Stanford GSB 26 Fund. Individual investors include faculty from Stanford, leaders in telecommunications, and AI experts from Anthropic.

Rather than depending on existing orchestration tools like Vapi and LiveKit, AethexAI has crafted its own compact model and orchestration layer from scratch, tailored to accommodate the regional English, French, and Arabic dialects prevalent in its target markets. This decision stems from the distinct operational challenges faced in these areas.

The company is also launching its platform, allowing enterprises to explore its technology and sign up for services, including APIs and SDKs for developers to experiment with its models.

Founded by Mariama Diallo and Ayooluwa Odemuyiwa, CEO Diallo previously held a position at Goldman Sachs before joining YC-backed ModelML in a product and growth capacity. CTO Odemuyiwa, an alumnus of Caltech, has experience with Meta and is currently studying at Stanford Business School before co-founding the firm. Their goal is to develop solutions for emerging markets while actively seeking investment opportunities.

Globally, there is intense competition among businesses to integrate AI tools for automating operations. However, success is not guaranteed. For example, an Egyptian call center automated a significant portion of its operations only to revert to previous methods due to poor performance, according to the founders. Many support centers in Africa continue to face challenges in recruiting engineers who can automate calls within budget limitations.

“The latency and jitter encountered with automated calls in this region were unacceptable. Had we chosen orchestration, we might have had to rely on large models hosted outside the region, leading to higher latency. We realized that to achieve success, we needed to utilize much smaller models and minimize latency at every step,” Odemuyiwa explained to TechCrunch, clarifying the reasoning behind the company’s development of its proprietary models and orchestration layer.

Typically, AI labs investing in cutting-edge models spend millions on training and data acquisition. AethexAI has discovered a novel approach. Instead of chasing the largest models, they found that smaller models would effectively tackle latency issues while maintaining accuracy. Their new Kora series features parameters ranging from 300 million to 1.7 billion, significantly smaller than conventional LLMs, fitting their objectives perfectly.

To train these models, AethexAI leveraged anonymized recordings from a call center partner. They also distributed hard drives to various radio stations across Africa to accumulate additional audio data. To further cut costs, they formed a contributor network of university students for data annotation and correct pronunciation of local names. As a result, the startup reports managing over 17,000 calls daily.

From a business perspective, AethexAI is dedicated to guiding clients new to voice AI through the process by offering on-site demonstrations and workshops to help them identify optimal automation use cases.

“We constantly remind our clients that we can’t meet every need at this moment. As a small company, we ask them to focus on one critical use case for our initial engagement,” Diallo noted.

While the startup is open to partnerships across various sectors, many of their current use cases revolve around calls related to debt collection, customer activation, and KYC — Know Your Customer processes, which are generally utilized by banks and telecommunications firms. They are hiring contract engineers for localized markets and collaborating with telecom providers to handle telephony for voice AI calls, stressing that one-size-fits-all solutions do not apply in this setting.

Walter Baddoo, co-founder and managing partner of 4DX Ventures, asserts that the challenges in the African and Middle Eastern markets are fundamentally different from those encountered by most voice AI companies.

“Companies in Africa and the Middle East manage about three times the call volume of their Western counterparts, as voice remains the primary channel for customer interactions,” he explained. “Existing systems were created with Western markets in mind, characterized by advanced GPU infrastructure, standardized English, and European speech patterns, along with enterprise workflows typical of the U.S. and Europe. Consequently, there are significant gaps for enterprises that require systems capable of handling dialects, code-switching, and informal speech while fitting into their existing telephony infrastructures and budget constraints.”

In summary, while firms like ElevenLabs, Deepgram, Sierra, and Cognigy are rapidly growing globally, the markets they are designed for and those they are entering may not always align. Startups like AethexAI are confident that these disparities—such as models customized for local dialects, on-the-ground partnerships, and region-specific infrastructure—represent a market opportunity that larger firms may lack the incentive or architecture to pursue.