A
We build
Ventures are conceived, engineered and taken to market in-house — from data pipeline to interface, on real infrastructure, with real users.
AI Venture Studio & Holding Company
Frontier AI, built to be owned — not rented.
Maluti Corp develops, operates and holds equity in AI ventures across South Africa and the United States. Nine products, engineered so that the model, the data and the decision stay inside our customers' walls.
09
Ventures developed
and held
and held
02
Operating markets
ZA · US
ZA · US
09
Sectors addressed
across the portfolio
across the portfolio
03
Regulatory regimes
engineered for
engineered for
01 / The thesis
Intelligence is becoming infrastructure.
Infrastructure should be owned.
The last decade of software taught organisations to rent their systems. The next decade will not be so forgiving. When a model decides who gets credit, which patient is at risk, or which footage a control room sees, the question of where that model runs and who can inspect it stops being an IT preference and becomes a matter of sovereignty.
Maluti Corp exists to build that generation of AI companies — and to hold them. We are not a consultancy that ships and leaves. We take equity, we operate, and we carry the products through to production in the markets we understand.
B
We hold
Maluti Corp retains equity across the portfolio rather than selling services. Long horizons, patient capital, compounding technical assets.
C
We operate
Shared architecture, shared guardrails, shared compliance posture. Each venture inherits the platform work the others have already paid for.
02 / Capability
Four things we do that most AI companies do not.
01
Sovereign AI
Every venture is designed to run where the customer's law says it must — on their own hardware, at the edge of their network, or air-gapped entirely. Behavioural Engine and Supervision ship as containerised stacks on customer hardware with no cloud dependency and no data egress. Legal AI can run fully local so client files never leave the box.
Built on open-weight models where it matters, so cost is compute — not per-analysis licensing to a vendor who can change the terms.
02
Bespoke AI on frontier research
We do not resell a general assistant with a new logo. Each product is built around one problem and one domain — multimodal fusion for poultry health, cross-camera person re-identification for audience behaviour, memo-grounded marking for matric exams, event-sourced graphs for revenue and org design.
Current-generation language and vision models are used deliberately and economically. AceirMatric marks both text and handwritten vision answers on a single unified model at roughly $0.00004 per mark — with no GPU to own.
03
Agent-native judgement layers
Systems of record — the CRM, the HRIS — describe an organisation as it is. Our agent-native platforms sit above them and treat the organisation as an object to be continuously optimised: goal-holding agents that audit, detect, simulate and propose, against objectives leadership sets.
They launch read-only over an existing system. No migration, no rip-out, no write-back in v1. Humans approve; agents prepare the case.
04
Verification, not vibes
The language model plans and extracts. Deterministic code grades the result. In Legal AI, every citation is checked for existence against real court records and every quote matched against the actual opinion text — with no model involved in the verdict, and a preflight gate that hard-blocks the pipeline on an unverified citation.
The same stance runs through the portfolio: momentum and coverage maths are deterministic, alerts must be corroborated by more than one signal, and every recommendation carries the evidence trail that produced it.
03 / Architecture
One stack, four layers, nine ventures.
Every Maluti product is assembled from the same layered architecture. It is the reason a nine-venture portfolio can be run without nine separate platform teams.
Layer 04Experience
The surface a human actually works on — exam room, control room, dashboard, attorney status card, approval queue. Designed so the consequential decision is always presented to a person.
Layer 03Judgement
Goal-holding agents and scoring engines: threat scores, momentum scores, readiness scores, structure audits, forecast models. This is where the product's opinion lives — and where it is versioned so accuracy can be measured after the fact.
Layer 02Verification
Deterministic guardrails between the model and the user: citation and quote checking, multimodal corroboration, memo grounding, preflight gates, PII masking and egress control. Nothing reaches the surface ungraded.
Layer 01Substrate
Compute and data, placed where the jurisdiction requires: customer hardware, edge devices beside the cameras, a regional VPS, or serverless inference. Event-sourced storage so any past state is reconstructable.
04 / Portfolio
Nine ventures. Two families.
Applied AI products built for a specific vertical, and agent-native or verification-led platforms that add a judgement layer on top of an existing system of record.
EducationLive
AceirMatric
AI-marked NSC exam prep grounded in roughly 972 real DBE past papers and official memos. Never invents a question, never hands over the answer.
aceirmatric.co.za→
FinanceCore built
ZarloPay
Cross-border cards, FX, remittance and wallet for travellers and SMEs — with AI layered in for fraud monitoring, KYC and expense intelligence.
ZarloTech (Pty) Ltd→
Financial inclusionIn development
microlendr
AI-underwritten micro-loans up to R300 for NSFAS, bursary and SRD-grant recipients, with every loan and repayment written to an immutable ledger.
Blockchain ledger→
AgricultureIn development
Poultry AI
Edge-first intelligence for commercial broiler and layer farms: health, weight, feed, environment and welfare, from cameras, microphones and sensors.
Edge / multimodal→
Audience analyticsIn development
Behavioural Engine
Self-hosted video analytics that fuses overhead and eye-level cameras to answer what no single camera can — who engaged, and who merely passed by.
Self-hosted→
SecurityIn development
Supervision
Real-time CCTV threat detection over standard RTSP. Three models in parallel produce one 0–10 threat score; the control room is notified in seconds.
On-prem / edge→
LegalIn development
Legal AI
Plaintiff-side research and evidence platform whose citations are verified against real court records — and whose pipeline blocks on a hallucinated one.
Single-tenant→
RevenueIn development
Agentic Revenue Engine
A judgement layer above the CRM. Treats pipeline velocity as the objective, forecasts from behaviour rather than stage labels, and proves accuracy before it ships.
Agent-native→
Organisational designIn development
Form
A judgement layer above the HRIS. Treats the organisation as a living object — audits spans and layers, models futures, and exports approved change plans.
Agent-native→
05 / Deployment
Two postures. Your choice, not ours.
Whether you need the elasticity of managed infrastructure or the strict isolation of sovereign hardware, the products are built to sit on either side of that line.
Posture A
Sovereign
Your hardware, your network, your jurisdiction. For regulated data, surveillance footage, privileged client files, and anywhere egress is not an option.
- Runs on customer-owned hardware, on-premise or at the edge
- Air-gapped and offline-tolerant operation
- Open-weight models — compute cost, no per-analysis licensing
- Single-tenant: one customer, one server, one database
- Local inference option so data never leaves the box
Posture B
Managed
Regional hosting with serverless or managed inference. For workloads where elasticity and cost-per-unit matter more than physical isolation.
- Regional hosting chosen for data-residency alignment
- Serverless inference — no GPU to own, cents at pilot scale
- Portable reference architecture across major clouds
- Tenant isolation and pseudonymised identifiers by default
- No training on customer data without explicit opt-in
Sovereign AIHuman in the loopEdge-firstPOPIAGDPRCCPAOpen-weight modelsEvent-sourcedDeterministic guardrailsNo data egressVerified citationsMultimodal fusion
Sovereign AIHuman in the loopEdge-firstPOPIAGDPRCCPAOpen-weight modelsEvent-sourcedDeterministic guardrailsNo data egressVerified citationsMultimodal fusion
Bring us the decision
you cannot afford to get wrong.
Pilot, licence, bespoke build or partnership — the conversation starts the same way. Tell us the operational bottleneck and where the data is allowed to live.