Cape Town startup Verascient has raised $1.2 million (R19.5 million) in an oversubscribed pre-seed round to build infrastructure that gives AI agents access to an organisation’s accumulated context — and it got there by abandoning the product it started with.
The company first built a hallucination detector for AI models. Its founders scrapped it after concluding that large language models were improving and hallucinating less, and rebuilt around enterprise deployment infrastructure instead.
That is an unusually direct judgment for a startup to make in public. Most companies do not retire a product because the problem it solved was shrinking. It also puts Verascient on the opposite side of a bet that plenty of others are still making — including the South African insurance sector, where Santam argued this month that AI hallucination is “a foreseeable engineering risk” rather than a diminishing one, with direct consequences for professional indemnity cover.
What it actually builds
At the centre of the platform is a temporal knowledge graph — a structure that captures not just what an organisation knows but when it knew it, who had access, and where the information came from.
“It builds and maintains a comprehensive understanding of an organisation, ensuring that information that was once scattered quickly becomes securely available to AI agents while preserving the history, permissions and provenance behind the information,” said Emile Ferreira, co-founder and chief technology officer.
The platform includes an agent-to-agent messaging protocol, background agents that run without user prompts, role-based access control and more than 1,000 integrations with existing business systems.
Verascient targets financial services, insurance and logistics — sectors where institutional knowledge is scattered across emails, documents, meetings, spreadsheets and internal systems, and where the value of an AI agent depends almost entirely on whether it can reach that material.
The round
Founder Collective led, having previously backed Uber, Airtable and Whoop. Additional capital came from Cambridge Enterprise, the University of Cambridge’s investment arm, Andrena Ventures under Gideon Valkin, and Summit Ventures. Angel investors Alan Knott-Craig and Shayne Mann also participated.
Ferreira taught himself to program at 12 and later worked as a developer at Replit, now valued at more than $9 billion. He built an on-device AI assistant while at school that reached over 200,000 users, and holds an MPhil in Advanced Computer Science from Cambridge. He co-authored a paper on enterprise AI ownership costs with the United Nations International Computing Centre and published energy efficiency research at the 2026 IEEE/ACM International Conference on AI Engineering.
Co-founder Keagan Stokoe was on the founding team at Fibertime, the South African pay-as-you-go fibre provider serving more than 1.5 million users a month, and later founded Purple Dorm, a consultancy helping organisations in South Africa and the UK deploy practical AI.
Engineers as product
Verascient does not simply ship software. It deploys its own engineers to work alongside customers on implementation.
“Ambitious leaders who see AI as a chance to reinvent how their company works will need people who understand both the technology and the realities of deploying it inside an organisation,” Stokoe said. “That’s why our model combines infrastructure with engineers working directly alongside our customers.”
The reasoning is sound on the evidence. MIT’s State of AI in Business 2025 found generative AI pilots failing at high rates for reasons of implementation rather than model capability. Cape Town consultancy Specno’s own 2026 report reached a similar conclusion about South African product teams — that the constraint is judgment and process, not engineering velocity.
The risk is structural. Forward-deployed engineering is how infrastructure companies win early enterprise contracts, and also how they end up as consultancies with headcount-linked revenue and consultancy margins. Palantir made it work; many have not. Whether Verascient can retire the services component as the platform matures is the question the model raises.
Where it sits
Verascient enters a Cape Town enterprise AI cluster that has been drawing capital steadily. Cue raised $5 million in July for autonomous customer service agents. Specno has been publishing on the product-to-revenue gap. Untapped AI is building voice automation for South African accents.
Verascient’s position is one layer beneath those — the context and memory infrastructure that agents in any of those products would need to work well inside a real company.
The capital will fund engineering hires, with the founders explicitly recruiting in South Africa. “We are looking for high-agency people with the curiosity, technical depth and ambition to build something meaningful to solve real-world problems,” Stokoe said.
