Nine ventures,
one engineering doctrine.

What learners get

• Exam room. Complete past papers rendered as printed, with figures, shared context blocks and per-subject answer formats: free text, a maths equation editor, real Accounting answer-book grids and a drawing canvas.
• AI marking rail (ACEIR). Per-component method and accuracy marks, a breakdown of what was earned versus lost, and memo-grounded guidance nudges.
• Adaptive smart practice. AI-composed practice sets and timed mock exams with NSC pacing and sitting rotation.
• APS simulator. Choose a career field and programme to see every institution offering it side by side — the full public university, UoT and TVET catalogue — with per-subject requirement checks.
• Exam Readiness score. Coverage × performance × exam practice, kept distinct from APS.

Other roles & commercial model

• A teacher and school console with cohort readiness, roster and per-learner detail.
• A parent view linked by a learner-issued code, with no answer text exposed.
• A Django-admin CMS for content review and approval.
• Schools purchase a licence with seats; a superadmin grants learner access. The licence-request flow is handled in-app, with payments manual by design for now.

Stack & economics

Django 6 and SQLite on a South African VPS (~R209/month, in a POPIA-friendly region), with nginx → gunicorn and git-driven deploys. AI runs serverless via Hugging Face → Qwen3.5-9B (DeepInfra) — a single unified model for both text and vision marking, at roughly $0.00004 per mark. A 100-learner pilot therefore costs single-digit dollars a month, with no GPU to own.

Status & open items

Approximately 20,000 approved questions are live across 10 text subjects. Maths and Maths Literacy are scans being transcribed sitting by sitting, and POPIA compliance (consent, export, delete) is built.

• POPIA data residency — inference currently leaves South Africa
• Online licence payments
• Maths and Maths Literacy scan backlog
• A live-AI re-review pass to enrich the bulk-approved corpus

What users get

• Cards. Create virtual cards for online use and order physical cards for in-store spending, with freezing, spending limits and other controls.
• Multi-currency. Exchange at real-time rates and convert and hold multiple currencies in one wallet.
• Payments and transfers. Domestic and international payments, instant transfers between ZarloPay users, and third-party wallets such as Apple Pay and Google Pay.
• Expense management. Track spending by category and date, set budgets, and generate expense reports and summaries.
• Onboarding and security. Registration by email or phone, KYC identity verification and multi-factor authentication.

Where AI fits

AI is being introduced where it adds the most leverage: continuous transaction monitoring for fraud and money-laundering detection, automated KYC identity verification from uploaded documents, and intelligent expense categorisation and reporting. Real-time notifications keep users informed of transactions and currency-rate changes.

How it works

• Micro-loans. Short-term advances capped at R300, deliberately sized to a student or grant-recipient budget.
• Eligible borrowers. NSFAS and bursary students and R370 SRD-grant recipients, verified against their funding or grant status at onboarding.
• Mobile-first flow. Application, credit decision, disbursement and repayment handled in-app, with repayment scheduled around funding and grant cycles.

Where AI fits

Because the target borrowers have thin or non-existent credit files, AI performs the underwriting from alternative signals — funding or grant status, transaction and repayment behaviour, and affordability indicators — to reach a responsible credit decision and set an appropriate limit within the R300 ceiling. The same models drive identity and fraud screening and predict repayment risk, so limits stay affordable rather than maximised.

Where blockchain fits

Every loan, its terms and every repayment are written to an immutable, auditable ledger, and smart contracts govern disbursement and repayment so agreed terms cannot be altered after the fact. Borrowers get a transparent record of what they owe and have repaid; the business gets a tamper-proof audit trail that supports both trust and regulatory reporting.

Responsible lending & compliance

Lending to students and grant recipients is sensitive and, in South Africa, tightly regulated. microlendr is positioned around affordability-first, responsible lending under the National Credit Act, registration with the National Credit Regulator, transparent pricing, and POPIA-grade handling of borrower and grant data. The low R300 ceiling and short terms act as built-in guardrails against over-indebtedness.

How it is built

The platform is edge-first: high-volume video and audio are processed near the poultry house, while events, features and selected evidence are transmitted onward. It stays multimodal (no high-impact alert rides on a single noisy signal), offline-tolerant (local monitoring and PLC safety continue through connectivity loss), and explainable (every alert shows its contributing signals, confidence and recommended next action). A reference implementation maps cleanly onto Azure IoT and AI services but is portable to other clouds or on-premises.

Operating boundary & safety

Poultry AI is explicitly an early-warning and decision-support system, not an autopilot. Disease confirmation remains a veterinary and laboratory process; feed and welfare changes stay within the farm's approved nutrition and equipment rules; and a PLC retains final deterministic control with hard limits, fallback schedules and manual override. In the South African context, deployments that capture identifiable people are handled under POPIA with a privacy impact assessment.

Differentiating capability

Its differentiating capability is fusion: linking the same physical person across both camera perspectives to answer questions no single camera can — for example, whether the visitors who engaged with a display were also the ones who dwelled longest. Results are explored through a web dashboard and a natural-language Query tab powered by a locally run Qwen model, so business users can ask questions in plain English and receive answers grounded strictly in the computed numbers.

Commercially relevant properties

• On your hardware. Runs entirely on the customer's own hardware via Docker (CPU or GPU), with no cloud dependency or data egress.
• Privacy-processed output. Produces privacy-processed video outputs alongside the raw annotated footage.
• Reusable event library. Maintains a searchable library of past events for reuse and comparison.

The problem

Security operations have far more cameras than human eyes. Incidents are missed in the moment, and reviewing footage after the fact is slow and manual. Supervision acts as a tireless first-pass filter that surfaces the small fraction of footage a person actually needs to see.

How it works

Three AI models analyse each feed in parallel: one identifies objects and people frame by frame, one interprets what the scene depicts, and one reads movement over time to recognise behaviour such as fighting, falling or crowd surges. Their outputs combine into a single 0–10 threat score. High scores trigger an immediate alert; lower scores are queued for human review.

What makes it defensible

• Audit trail. Every model call logs prompt version, tokens, latency and cost, attributable to a specific matter and reproducible after the fact.
• Privilege and privacy. Matter-level privilege tagging, PII masking, and egress and redaction controls; the model can run fully local (Ollama), so client data never leaves the box.
• Cost visibility. Per-matter spend is tracked live and surfaced on an attorney status card that answers what is next, what is blocking, and what it has cost.

Deployment

Deployment is single-tenant — one firm, one server, one database, one AI bill — built on Postgres, a FastAPI web UI, and TLS via Caddy, installable on a VPS with a single script. A desktop packaging path and magic-link (passwordless) login are also available; the platform is licensed under AGPL-3.0.

As stated prominently in the repository itself, output is not legal advice and requires attorney review before it is relied upon or filed.

The agents

• Hygiene Agent. Makes the pipeline picture true — detects zombie deals, stale stages and duplicates, and measures coverage against target.
• Deal Agent. Maximises win probability per opportunity — momentum scoring, stall detection from activity decay, and drafted next-best actions that also save reps time.
• Forecast Agent. Produces a forecast leadership can defend — probabilistic, with confidence intervals, projected from behavioural signals and shipped only behind a backtesting harness that proves accuracy first.
• Motion Agent. Improves the sales motion itself by analysing won and lost patterns across the event history.

What makes it defensible

An event-sourced revenue graph treats activity as evidence and CRM stage fields as mere claims; every recommendation carries an evidence trail traceable to the signals that produced it; and forecasts are versioned so accuracy is measurable after the fact.

The deterministic-versus-LLM split is deliberate — momentum and coverage maths are deterministic, while stakeholder-role inference and action drafting use language models. In v1, no agent writes back to the CRM or communicates externally; humans approve and act.

How it works

• Objective-driven. Leaders set the targets the agents optimise against — spans of control, layer limits, cost envelopes, capability coverage — so the product stays industry- and location-agnostic by design.
• Simulation. Draft org states are first-class: multiple proposed futures can be modelled in parallel, compared on predicted cost, span, layer and capability metrics, and promoted into an approved change plan.
• Time travel. An event-sourced graph makes any past org state reconstructable and every proposal diffable against any state.
• Evidence and approval. Every recommendation carries the data, reasoning and confidence behind it, and human approval gates are part of the domain model, not a UI afterthought.

The agents & privacy

An Integrity Agent makes ingested data trustworthy; a Structure Agent audits spans, layers, duplication and cost concentration and drafts restructures; a Capability Agent closes the gap between skills required and skills held; and a Scenario Agent models workforce futures against strategic inputs such as “what does this org look like at 2× revenue?”

Employee data is treated to the strictest common denominator across GDPR, POPIA and CCPA — data minimisation, an option to run on pseudonymised identifiers rather than names, tenant isolation, regional hosting, and no training on customer data without explicit opt-in.