Research

Marthinus G.T. Ferreira

Chief Technology Officer — AI infrastructure, evaluation and governance

Viable Core Business · Johannesburg & Pretoria, South Africa · CISSP #73655

tommy@vcb-ai.online

I work on the measurement problems that sit underneath enterprise AI: how evaluation procedures behave when their statistical assumptions fail, and what that means for organisations that have to certify a system before deploying it.

Thirty years in infrastructure across banking, telecommunications and regulated enterprise in Africa, latterly building AI platforms and legal-technology products. The research interest follows directly from the practice: production systems are validated by procedures whose guarantees are conditional, and the conditions usually go unmeasured.

Publications

Consensus Gates Hide Rare Defects Only Inside a Narrow Correlation Window
Under peer review, 2026 · An exact suppression law for threshold group testing

A closed form for how far a majority-of-k gate understates a rare defect rate under correlated judgements, with the asymptotics of the half-suppression threshold. The concealment window closes quadratically in the number of judges, so adding voters raises peak suppression while shrinking the regime in which it is real. Includes a measurement on 91,504 human-annotated items showing that real annotator panels sit far outside that window.

Download preprint (PDF) → Verification corpus — 32 claims, 34 theorems →

Suppression factor against intra-class correlation, log-log, for gate sizes k=3 to 11 at a 5% defect rate. Markers show the half-suppression threshold.
Suppression against correlation at a 5% defect rate. A nine-judge gate begins 1,505× suppressive and is halved by ρ = 0.38%, while a three-judge gate survives to 5%. Markers sit at the half-suppression threshold.
Suppression ratio against the single parameter a = q(1-rho)/rho, showing all gate sizes collapsing onto one curve each.
The one-parameter collapse: every combination of defect rate and correlation sharing a value of a = q(1−ρ)/ρ lands on the same point. Where each curve crosses the dashed line is the half-suppression root.

All eight figures, interactive →

Email Exposed: Fundamentally Flawed
International Journal of Science, Technology and Society, 2025 · ISSN 2330-7420 · Manuscript #1831765

Cryptographic analysis of SMTP vulnerabilities, with an evaluation of SPF, DKIM, DMARC and DNSSEC, blockchain-based authentication, and AI-assisted threat detection. Peer-reviewed.

Unravelling the Illusion of Email Security
Preprint, April 2024

doi:10.13140/RG.2.2.14664.53764

Technical reports

Tinman Training Factory — Run Report
Internal report, 31 July 2026 · Sixteen training jobs, five days, one shipped model

A complete record of the Tinman fine-tuning pipeline as it actually ran, 26–30 July 2026 — every job, every loss curve. Two model families were trained: Qwen3-8B, which shipped as the model behind Tinman@Labour AI, and Gemma-4-E4B, evaluated and set aside.

Full run report →

Interests

Threshold group testing Evaluation under correlated error Model validation & assurance Sovereign AI infrastructure GPU compute at scale Security governance Legal AI & RegTech

Selected work

  • PRIME — co-architect of a model-agnostic sovereign enterprise AI platform: persistent intelligence database, multi-model orchestration, vendor independence with hot-swap, POPIA-aligned governance.
  • LIANELA VOICE — a legal reasoning engine fine-tuned end-to-end on a curated South African legal corpus, built from corpus curation through to production deployment and reviewed by High Court advocates.
  • Tinman Labour Law model training — trained a South African labour-law model end to end: sixteen fine-tuning jobs across two model families, evaluation, and the shipped model behind Tinman@Labour AI.
  • Curated legal corpus build — a hand-selected South African legal corpus curated by High Court advocates rather than scraped, and published with a claim-level audit that records the unconfirmed claims alongside the confirmed ones.
  • GPU cluster architecture — designed and delivered a large on-premises NVIDIA GPU cluster inside a regulated South African financial institution: InfiniBand interconnect, high-density compute, air-gapped governance.

Background

Chief Technology Officer at Viable Core Business since 2025, leading AI platform development and advisory delivery. Previously in enterprise architecture and AI infrastructure roles at Altron Digital Business, Data Sciences Corporation, First Technology National and Hitachi Data Systems, beginning in network and mainframe operations at Medihelp in 1994.

CISSP since 2006 (#73655). NVIDIA GPU Genius and DGX certified. Professional member of IITPSA, ISC² and ICFP. BCom, University of South Africa.

Enquiries about the research, or about model-validation and evaluation-assurance work, are welcome by email. Code and data-preparation scripts supporting the preprint are available on request.