Quelin Zammit

I build and operate unattended automation for money-touching workflows, the kind that has to fail closed rather than guess.

For sixteen years I have been owner-operator and sole engineer at QPARTZ Ltd in Malta: I do the architecture, I ship it, and I am the one called when it breaks. The platform is React and TypeScript on PostgreSQL, 36 tables governed by 79 row-level-security policies, 234 shipped migrations and 21 serverless edge functions, running live bookings, payments and back-office operations, with more than 150 production automations across finance, operations, compliance and marketing. The payment pipeline is server-authoritative, with HMAC webhook verification, replay protection, business-level idempotency and automated orphaned-payment detection.

Since 2022, under non-disclosure, I have built an execution control plane that drives a live retail banking portal over Chrome DevTools Protocol and places real orders unsupervised, behind five independent fail-closed gates: pre-commit verification of instrument and exact quantity, a guard that refuses to act when it cannot read the terminal control, a three-strike circuit breaker, a per-transaction ceiling, and idempotency keys. Beside it sits a reconciliation engine that classifies drift into six typed, severity-tagged categories rather than merely flagging it.

My AI work is about the harness, not the model. QUANTZ V3 is a Python platform running a full research cycle unattended on a dedicated 32-core machine, with an LLM research brain built on a hard information barrier: four roles holding four disjoint information sets, so the component proposing parameter ranges structurally cannot observe out-of-sample results. Versioned prompts, schema-constrained output, numeric guardrails, per-window spend budgets, and an audit trail of every payload and decision. Every numeric threshold lives in one versioned manifest, enforced by a test that fails the build if any literal from it appears in the guardrail source. A separate pipeline carries machine-written artefacts through roughly 2,000 conformance checks and six months of mandatory forward validation before anything runs near real capital.

Governance is designed in, not bolted on: a human-owned safety core with evaluation-gated promotion mapped to EU AI Act Article 14, and deterministic PII redaction validated by Luhn, IBAN mod-97 and VIN checksums.

Payments have been the through-line since 2015, when I worked platform-side on card payments software at RS2, debugging against production banking systems. Eighteen years writing software, sixteen running production systems, across payments, trading, compliance, support and systems administration.

Maltese national and EU citizen, so no work authorisation is required anywhere in the EU. I am looking for senior process automation or AI systems engineering work in fintech or financial services, remote across Europe.