Deploying global voice AI in African clinical settings exposes an immediate structural flaw: the models cannot parse local accents, code-switching, or high-noise environments. The result is systemic failure, translating to corrupted data and compounding clinical delays.
Tobi Olatunji, a physician-turned-machine-learning-scientist, founded Intron to build a localized speech infrastructure capable of handling Africa’s phonetic realities. What began as a solution to critical bottlenecks in medical documentation has since expanded; today, Intron’s flagship model, Sahara, powers voice AI for courts, call centres, and banks across the continent — trained on over 3.5 million audio clips from 40,000 speakers.
In this interview with Launch Base Africa, Olatunji outlines the mechanics of building proprietary datasets, the unit economics of their infrastructure-first pivot, and why the real bottleneck for African AI is post-signature change management.
1. What does Intron do, and what specific, on-the-ground reality convinced you that this infrastructure was non-negotiable?
Intron builds voice AI and speech infrastructure for Africa — the accents, the code-switching, the names, the medical terms. Sahara is our flagship model.
The turning point wasn’t a single event, but observing the baseline error rates in high-stress clinical environments. In a typical West African ward, you have severe ambient noise — phones, overlapping conversations, intersecting patient needs. When a physician is overloaded, manual documentation inevitably drops critical data points, like specific biomarkers. If the initial data capture fails, the entire clinical record downstream is compromised. We realised that any voice AI deployed here couldn’t just optimize for speed; it had to be engineered specifically for high-noise, high-pressure environments, capturing data accurately regardless of the speaker’s accent.
2. What is the founding team like, and how has the partnership grown
Kunle (Olakunle Asekun) brought enterprise scaling experience from GE, KPMG, and IE Business School. My background bridged medicine and technical execution, moving from clinical practice to NLP research and machine learning at AWS. I wrote my first line of code in 2018. We engineered the founding team to explicitly cover both the commercial and technical vectors required to scale AI infrastructure. From that initial pairing, we have scaled to a highly specialized team distributed across the core markets we serve.
3. Intron started in health and now serves courts, call centres, and banks. Was the expansion planned or pulled by the market?
The market pulled us, in ways I did not fully anticipate. We started in healthcare and kept running into the same problem: existing speech recognition tools could not handle African accents or the specific vocabularies of the domains we were working in. So we built our own.
What we did not expect was that the same capability solved the same structural problem in entirely different sectors. The courts are a clear example. Nigeria’s legal system had long been trapped in the analogue age. Ogun State was spending four hours on cases that now conclude in two or three. The Chief Registrar told us: “My Lord no longer writes during proceedings. He listens. He focuses.” That happened because we had built for the hardest acoustic conditions first, and the underlying model transferred. The roadmap shifted from healthcare-first to infrastructure-first.
4. What has been the single most unexpected hurdle deploying across different African sectors?
The assumption we had to kill earliest was that the technical problem was the hardest one. It is not.
One of our earliest lessons was learning to design for the first-time user: two buttons, record and stop, copy and clear. That discipline came from watching good technology get resisted — not because it did not work, but because it disrupted workflows institutions had organised themselves around for years. You can deploy a model that performs well on a benchmark and still lose to a paper form. That changed how we think about a signed contract. Signature is the beginning of the work, not the end of it. What follows is change management, training, identifying internal champions, and iterating on what real usage data shows you.
5. When global tech giants eventually prioritise African accent diversity, what is your actual moat?
This question assumes the gap is primarily a matter of attention. It underestimates the difficulty of the data acquisition.
Sahara’s model performance — recognising over 500 African accents at above 92% accuracy — is built on domain-specific data pipelines. You cannot scrape years of clinical voice data from Nigerian, Ugandan, and Rwandan health systems off the internet. It requires establishing complex, localized consent frameworks and fine-tuning models for highly specific workflows. That proprietary asset cannot be replicated remotely. Furthermore, enterprise infrastructure requires localized accountability. A global API failing in a Nairobi hospital offers no immediate engineering support. Our physical and operational presence in these markets is our actual moat.
6. What is the most underestimated bottleneck — the one that only reveals itself after you are already in the market?
Data quality at scale. Not quantity. Quality.
What people consistently underestimate is how painstaking it is to collect data that is genuinely representative: accents across regions, real acoustic environments with background noise and overlapping speakers, and the specific vocabularies of the domains you are deploying in. Healthcare speech sounds nothing like legal speech. We crossed one million audio samples from 13 African countries before training at scale. Startups that bootstrap on publicly available datasets find out the hard way: the model works in the demo and breaks down the moment real users put it under load in the field. Data infrastructure consumed more of our early budget than anything else. It is the single investment I would make again without hesitation.
7. What conventional fundraising advice have you deliberately ignored, and what happened?
The standard playbook for African tech is to aggressively maximize the round size and optimize for marquee names. We opted to structure our cap table around investors who viewed African infrastructure as a foundational asset class, rather than a speculative portfolio bet.
When market conditions tighten, the difference in board-level support is stark. Securing partners like NVIDIA, the Gates Foundation, and Google Research was a strategic choice to align our capital with our actual data and compute needs, rather than artificially inflating our valuation or misaligning our growth expectations.
8. You last raised in 2024. Where does fundraising stand today?
After the $1.6 million pre-seed round in 2024, we expanded our cloud-native and on-premise deployment capabilities and grew the engineering, sales, and research teams. We are currently focused on finding the right capital partner for the next phase — someone who understands African voice AI as a foundational infrastructure layer for the continent’s digital economy. We are taking our time with the decision.
9. What is the most humbling technical failure from the early days, and how did you correct it?
Accepting that the first iteration of our system did not solve the unit economics of time. The first doctors who tried the speech-to-text app during the pandemic took about 45 minutes to complete their notes — significantly slower than writing by hand.
The humbling part was returning to those early users, acknowledging the performance gap, and mapping out their exact data-entry friction points. Their feedback fundamentally changed how we optimized the latency and user interface of the current build.
10. What has kept you building at Intron through the tough stretches?
The feedback from the field is hard to walk away from. A doctor at EHA Clinics in Kano told us her documentation time dropped from 30 minutes to less than 10. Twenty minutes per patient is significant when you are seeing a full ward. What she described was not satisfaction with a product. It was a practical change in how her day worked.
I came to this because the problem I saw in practice was too urgent to leave alone. That has not changed. When the work is connected to something that concrete, the difficult stretches are easier to get through.
11. What is the one avoidable mistake that most first-time founders will repeat anyway?
Failing to distinguish between operational noise and structural traction.
In African markets, you are frequently presented with adjacent sector opportunities. It is easy to fill a calendar with partnership talks and assume the company is scaling, when in reality, the focus is just dispersing. Protecting our core data work required strict resource allocation — which meant passing on immediate, superficial revenue opportunities in secondary verticals to solve the harder, fundamental infrastructure problems first. The discipline is in being honest about what has actually converted versus what is just motion.