AI Meets the African Bush

From counting cattle to a conservation platform: how the idea started

The project began inside Chisl's consultancy work, not as a conservation mission. A friend approached the team about using AI and drones to verify agricultural assets so that corporates and farmers could secure funding against them. That agricultural verification work grew into Veriphy, an asset verification platform now used by Nedbank, Standard Bank and one or two of the big audit firms, and applied in the United States as well.

Kellerman and Strauss took it on partly because it was a challenge their engineers wanted. "We didn't care whether it worked a lot," Kellerman said. Much of the early work happened after hours, on weekends and in downtime between paying corporate projects. The livestock counting proved successful at scale, at one stage counting around 6 million livestock a year, but Strauss was candid that cattle alone would "probably not have the ultimate scale" the founders wanted. Conservation, a large industry with money flowing in from both commercial and CSI sources, offered that scale, and Kellerman was the one who pushed the idea.

Why satellites could not do the job

Satellite imagery was the obvious starting point, and it was the first thing to fail. Kellerman explained that satellite image quality typically sits between 15 and 30 centimetres GSD, or ground sample distance, meaning each pixel covers that much ground, and the lower the number the better the detail. That resolution was not good enough. Worse, a straight-down image, whether from satellite or from a drone pointed vertically, only ever identifies a fraction of the animals present. On current results from a straight-down view alone, Kellerman said, the system would pick up less than 20 to 25 per cent of individuals.

The answer was oblique imaging. The team moved to cameras with five lenses shooting at angles, capturing north, south, east, west and straight down, so that as the drone flew its grid it built up multiple views of the same ground, even partially under tree canopy. That solved the visibility problem but created a new one: the same animal now appeared in up to 20 images, so the hard engineering shifted to de-duplication rather than detection alone.

Two years, R800,000 drones and 45 degrees in the bush

Getting the models right took about two years, and getting them into the field was harsher still. The team started with DJI quadcopters, which they still use purely as image-capture devices rather than for any onboard AI, then invested in German Quantum Systems VTOL fixed-wing drones, the aircraft seen taking off in the project video. Flying at around 90 metres, the fixed-wing craft cover far more ground than a quadcopter.

The first big reserve deployment was a brutal introduction. The reserve ran to 45,000 hectares, the drones cost in the region of R800,000 each, and the operation demanded full SACAA certification and qualified pilots. The team hit 45 degrees Celsius and picked two of the windiest days of the year, with winds of 14 to 16 metres per second, upwards of 60 kilometres an hour. Around herds of buffalo and elephant, the pilots had to worry about the drone coming in to land at all.

Data was its own frontier. Across the campaign the drones gathered more than 40 terabytes of data and more than 35 million images, some 7.8 billion pixels, all of which had to be pulled off the aircraft twice a day onto encrypted hard drives in the bush, on solar power with no mains electricity. "We became experts around data storage," Kellerman said.

This episode was sponsored by

Silicon Overdrive Founders Series

Visit Silicon Overdrive↗

Old-school detection models, AWS H200 GPUs and an evidence-based census

Once the raw data was back in Johannesburg, the analysis needed serious compute, and AWS provided it. AWS, already a Chisl partner on large-bank work, donated 8 Nvidia H200 GPUs, more than a terabyte of virtual RAM, which Kellerman noted is more processing power than much of corporate South Africa has. That in turn forced the team to rebuild their platform and data pipelines so they could feed the GPUs fast enough to avoid bottlenecks.

Notably, the heavy lifting is not done by large language models. Kellerman was clear that the identification uses "old school detection models" such as YOLO, RT-DETR and Faster R-CNN, deployed in stages depending on the use case, with YOLO fast but less accurate and the transformer-based RT models strong on small objects. Large language models, run through Bedrock, are reserved for validating data and helping users query the imagery, not for the detection itself.

What sets the method apart, Kellerman argued, is that it is evidence-based rather than statistical. Traditional counts are smart and innovative but not always backed by proof. Here, every identified animal comes with a photograph and a GPS-referenced location. On enclosed reserves where owners already knew their numbers, the team could test itself against a known answer and, in Kellerman's phrase, there was "nowhere you can hide." The results proved as accurate or more accurate than traditional counts, with feedback coming from the reserve owners themselves at Sabi Sands, Timbavati, Kwandwe and Kariega.

Where Veriphy goes next: from census to ecological platform

The founders now see Veriphy as a holistic platform rather than a once-a-year headcount. The same flight data can map roads, water bodies, trees, erosion and the impact of elephants on vegetation, and can be re-run against new models when something new becomes of interest, such as vulture nests. Strauss pointed to carbon credit and green bond applications, where funders need comfort that money is being applied as intended, and to asset financing in agriculture.

The near-term target is scale: one million hectares, done accurately. Kellerman was blunt that doing a million hectares is easy but doing it accurately is not. The team is also building its own drones and experimenting with onboard real-time AI, and sees applications for the underlying computer vision well beyond conservation, including financial services and healthcare.

[CHAPTERS]

Why not just use satellites to count wildlife?

Satellites are not precise enough. Kellerman said satellite imagery typically sits at 15 to 30 centimetres GSD, which is not good enough to reliably identify individual animals, and a straight-down view captures less than 20 to 25 per cent of individuals. Veriphy instead uses drones flying at around 90 metres with five angled lenses, building multiple overlapping views of the same ground so that partially hidden animals are still detected.

Can this technology be used for anti-poaching?

Yes, in principle, but the founders are deliberately keeping it separate. Asked directly by an audience member about predictive anti-poaching, both Kellerman and Strauss said the capability clearly exists but sits in a ring-fenced, separate entity because it carries complications closer to defence and security. Veriphy itself stays focused on conservation and ecological surveys, which capture images for processing afterwards rather than the real-time analysis anti-poaching would require.

What stops a competitor from copying this?

The moat is the data, not the code. Strauss acknowledged that two years ago the software itself was a moat, but that has largely disappeared with modern coding tools. What is hard to replicate is gathering data at this scale, which is neither cheap nor easy, along with the trust and credibility needed to operate in a conservation industry that is not simple to break into. That combination, Strauss said, is not what keeps them awake at night.