Systems that earn their place in production.

A reproducible pipeline that downloads official commission-of-inquiry transcripts, speaker-chunks the text, embeds every passage into Qdrant for semantic search, and builds an evidence-aware knowledge graph in Neo4j — served through a read-only FastAPI + React surface that walks from a search hit to its chunk, its mentions, and each claim's full provenance.

The data model enforces the principle that a mention is not a claim, a claim is not a finding, and a finding is not a fact. First corpus: the Madlanga Commission — 106 hearing days, 18,192 pages, 81,712 speaker turns, ~3.65M words, and roughly 49,000 LLM-extracted claims, every one traceable to its PDF page.

An AI scribe that listens to the consultation, structures the clinical note, and hands the practitioner back the time they previously spent typing. Positioned as documentation support — explicitly not a medical device — and built POPIA-compliant by design, with the practice as responsible party and MedVoice as operator.

Gauteng-first go-to-market with founding-pilot pricing for the first ten GP practices, backed by an ops repo where a coordinated set of AI agents — researcher, sales, marketing, engineering — run the company around the product from shared state.

A scheduled pipeline that discovers and crawls South African news and public sources, uses Claude for structured incident extraction, geocodes against a South African gazetteer, and assembles a Neo4j knowledge graph of perpetrators, victims, and routes — surfaced through a React dashboard with network and route analysis.

Built for the counter-trafficking domain KonteksWerke has served for years across face matching, advert analysis, and case-dispatch tooling for NGOs working on the ground. Alerting is wired in so a degraded pipeline announces itself instead of going quietly dormant.

A plugin-based OSINT platform for extracting, enriching, and analysing individual and network details from public web sources. Playwright collection with real-time progress streaming, phone-number intelligence, and Cosmograph network visualisation over an async FastAPI + React stack.

The research lineage runs deep: advert clustering and ranking, suspicion scoring, social-network reconstruction, and entity graphs built from classified-ad ecosystems — capability that feeds directly into the counter-trafficking and due-diligence work.

A pipeline that parses PDF bank statements into transaction rows, normalises descriptions and derives merchants, classifies every transaction against a maintained pattern library, and stores the result in SQLite as the source of truth — with monthly pivots, budget summaries, tax exports, and a signed ledger on the way out.

Classification patterns live in a Google Sheet that non-technical users can edit; the pipeline pulls them in, applies them, and produces an HTML review report. A monthly runbook makes the whole cadence repeatable rather than heroic.

A research and execution platform for systematic market timing, co-founded with a quantitative trader. The product combines macro signal extraction, market-profile tensor representations, and a reasoning layer that surfaces trade hypotheses with their underlying evidence trail intact.

Stage A business plan complete; Stage B in progress, covering MVP scope, technical architecture, AI system design, data strategy, and the trust and governance framework required for institutional use.

A progressive web app for school communications and fee collection, with a modular architecture that extends into adjacent operational domains. Phase 1 covers SMS and in-app messaging, Ozow and Paystack payments, Clickatell OTP, and QR / NFC-based learner cards.

A separate Clinic module is in design for the school nursing-care workflow, with accounting integrations researched and specified for Sage, Xero, and Odoo, over a PostgreSQL backend with Cloudflare R2 storage.