Full-stack engineer — TypeScript · Python · PostgreSQL · AI systems

I build software where being wrong has consequences, and I test it like it matters. Twenty-one years leading technical teams in safety-critical aviation across four countries taught me that; I now apply the same discipline to type systems, test suites, and reviewable code.

📍 Kuwait → relocating to Japan · 🇯🇵 residency comes through my wife, a permanent resident — no employer sponsorship required (spouse status held previously) 🌏 Open to remote worldwide, Japan-based, or hybrid 💬 English (native) · Japanese (studying)

Start here: zerofayyz-fintech.vercel.app — a deployed payments platform you can click through in three minutes, with the source beside it.

Gabriel — a desktop-class AI workstation platform, built solo. (private repo)

A single React interface over conversational AI, local image generation, image-to-3D mesh conversion, voice synthesis, an animation and storyboard studio, Blender and Unity integration, and a system-health lane that can inspect and repair the machine it runs on.

Every operation that changes state writes an auditable record. A system that acts on a real machine should be verifiable, not trusted — that principle shaped the whole architecture.

A cloud payments and operations platform — deployed, monitored, and reviewable. → zerofayyz-fintech.vercel.app · Stripe sandbox only; no real funds move, so everything is safe to click.

One Fastify/TypeScript API on PostgreSQL, consumed unmodified by three independent frontends — Next.js, Vue 3 and Svelte 5 — each validating every response against one shared Zod contract at the network boundary. Yen-denominated and bilingual, with translations enforced by the type system: a missing string is a compile error.

• Idempotency lives in a database constraint, not application branching, so a duplicate webhook is refused under concurrency and across restarts.
• An append-only audit log the application itself cannot rewrite — enforced by a database trigger, not by convention.
• Row-level security in a request lane: user-serving reads adopt a low-privilege database role per transaction, so PostgreSQL's own policies — not WHERE clauses — decide which rows exist. Proven by tests that SELECT with no per-user filter.
• Passwordless sign-in that stores no credential: the token in a magic link exists only in the email; the database keeps a SHA-256 hash of it. Single use is one atomic UPDATE, so two clicks racing on the same link produce exactly one session, and expiry is decided in SQL rather than by the application's clock.
• Four-eyes refunds: the requester cannot approve their own request, enforced by the API and by a CHECK constraint, and the ledger moves only on Stripe's signed event.
• An independent reconciler in Go, deliberately a separate process — a checker that shares code with the thing it checks agrees with its bugs.
• 389 automated tests across ten suites, all gated in ten CI jobs, including thirteen Cucumber scenarios that state the payment rules in plain language and execute.
• A durable job queue in the same database that holds the ledger — atomic claims over FOR UPDATE SKIP LOCKED, lease-based crash recovery, capped backoff and dead-lettering, surfaced to operators at /admin/jobs. The guarantee is at-least-once and is written down as such: exactly-once does not survive a channel that can die between doing the work and acknowledging it.
• Monitored in production on a schedule by a 30-check suite whose webhook probe signs a real event: a health endpoint reports that a signing secret is present, never that it is correct. I broke the secret on purpose to watch the alert fire, and restored it.
• A QA surface an AI agent can drive over the Model Context Protocol — six tools, with a protocol handshake check in CI. Its first human-driven run found a live defect.

The defect worth citing: the webhook handler had nineteen passing unit tests and had never once worked. Every test stubbed the database, so none executed the real SQL — and the SQL was invalid. The first integration test against real PostgreSQL found it immediately. A test suite has a shape, and defects collect where that shape does not reach.

Twenty decision records explain the trade-offs, including the ones still being carried.

A Manifest V3 browser extension that watches health endpoints from the toolbar — badge shows failures, one notification when an endpoint goes down and one when it recovers, never a repeat.

It ships watching the payments platform above, making it the fourth independent consumer of the same API — and the first from outside that repository, which is the strongest test of the contract claim. MV3 done properly: the service worker owns no state (Chrome kills it when idle), and host access is requested per origin at runtime, never as a blanket grab at install. The test most extension repos skip: Playwright loads the built extension into Chromium and proves the worker registers, its alarm exists, and both pages render.

→ receipt-portal-one.vercel.app

A customer's own receipts, built entirely on Supabase — Auth, row-level security policies, Storage, and one Edge Function on Deno that mirrors payments from the platform above. It is the fifth independent consumer of the same API, and it exists to enforce the same rule the platform enforces by hand — a customer sees only their own rows — the other way round, so the two row-level-security models can be compared honestly. That comparison is ADR 17 in the platform repo.

• Ten isolation tests against the live project: two signed-in customers, queries with no per-customer filter, rows absent because a policy refused them. Writes refused, other people's files refused, other people's signed URLs refused.
• The Edge Function holds the only privileged key at runtime. It authenticates its caller with a shared secret compared in constant time, answers identically for a mailbox it has never seen so it cannot enumerate accounts, and is idempotent — ninety-three rows on the first run, ninety-three on the second.

The trap worth knowing: with "automatically expose new tables" off — which Supabase recommends — a new table has no permissions for anyone, including the service role. Default-deny reaching the privileged lane; correct, and it has to be granted back on purpose.

A two-phase fair scan over a grouped document corpus, where group order never decides results.

Extracted from Gabriel's retrieval layer after I found a defect that made search silently answer from the wrong source — no error, no empty result, just a confident wrong answer, with a whole category of documents quietly unreachable.

The repo keeps the original broken implementation next to the fix and runs both on the same corpus, so you can watch it fail and then pass:

BEFORE  ordered scan with an early break
  found the answer : NO
  -> answered confidently from the wrong engine's documentation.
     No error. No empty result. Just wrong.

AFTER   two-phase fair scan
  found the answer : yes

Group order swapped -> identical results: yes

10 hermetic tests · zero dependencies · CI on Python 3.9 / 3.11 / 3.13

Languages TypeScript · Python · Go · SQL · C# · JavaScript · Bash

Frontend React 19 / Next.js · Vue 3 (Composition API + Pinia) · Svelte 5 (runes) · Vite · strict-mode TypeScript · runtime-validated shared contracts (Zod) · i18n enforced by the type system

Backend Node + Fastify · Python + FastAPI · PostgreSQL · REST API design · schema validation · webhook signature verification & idempotency · sessions, roles and rate limiting · background workers

Testing & QA Playwright (end-to-end, visual, accessibility) · Vitest · pytest · node --test · Cucumber / Gherkin · integration tests against a real database · axe-core in CI · ESLint 9 + typescript-eslint · tsc -b as a build gate

Infrastructure GitHub Actions CI/CD · Docker · Kubernetes manifests · Render / Vercel / Neon · database migrations · structured logging · error tracking · hourly production smoke monitoring

AI/ML Local LLM inference & model routing · RAG · embeddings · evaluation gates · MCP tool servers

Generative media & 3D SDXL / FLUX · image-to-3D mesh conversion · TTS voice cloning · Blender + bpy automation · Unity (C#) · Unreal

Platforms macOS · Linux · Unix-like · Windows

Alongside the engineering I build content tooling — SDXL and FLUX image generation running locally, image-to-3D mesh conversion, TTS voice synthesis, and two complete character rigs authored in Blender with Python (bpy) automation I wrote: weight transfer, a repeatable rig-transplant procedure that retargets a validated armature onto new meshes, 16 animation clips, and LOD-tiered FBX export. Unity and Unreal scene work, plus a 2.5D mobile platformer in progress.

The tooling exists because doing it by hand didn't scale — which is the same reason I write any tool.

B.S. Information Technology · University of Phoenix Advanced Cyber Security Certificate (Undergraduate, awarded with honor)

Before software: 21 years in aircraft maintenance and program leadership — U.S. Air Force, Lockheed Martin, DynCorp, Zenetex, AAR.

Poland — program systems owner for an 80-person USAFE F-16 depot. Authored the program's Performance Work Statement from the governing Air Force Instruction, then built the systems that satisfied it: qualification and training records from nothing, a leave-forecasting tool that surfaced manning shortfalls weeks early, and a digitised onboarding pipeline covering security-clearance investigation, fingerprint submission and expense reporting. The records those systems produced held up under a Defense Contract Management Agency review.

Japan — alternate site lead at MCAS Iwakuni, leading a team of eight on a C-130J programme, and built that site's training and qualification tracking from the ground up.

Oman — built the Microsoft Project system the Kuwait Air Force F-16 phase-inspection programme was planned and run from: work breakdown, dependencies, resource loading, schedule baselines. Also taught F-16 systems to non-native English speakers.

Work governed entirely by technical data, quality gates, and traceable records.

U.S. Air Force veteran · U.S. Secret clearance (renewed 2023)

marcel.gilbert.dev@gmail.com

Open to software engineering roles — remote worldwide or Japan-based.