I spent twenty years running restaurant and hotel operations. Then I built a fintech platform, put it in production, and people started paying for it.

PropDesk2026 —founder, architect, sole engineer

A control room for prop-firm traders

PropDesk routes a trader's own signals, normalises execution across brokers, enforces risk rules before orders are sent, and keeps every account and audit trail in one place. I designed it, built it, and I run it — the deploys, the monitoring and the 3 a.m. incidents.

Backend

Python, FastAPI, SQLAlchemy 2.0 over PostgreSQL with TimescaleDB and Redis. 21 modules, 161 versioned migrations.

Frontend

React 18 + TypeScript, Vite, Tailwind, TanStack Query, Zustand — a configurable widget workspace. Tauri desktop build.

Integrations

Seven broker and market-data providers, written from vendor protocols: TopstepX, Rithmic, Tradovate, cTrader, MetaTrader 5, dxFeed, MyFundedPerps. REST, WebSocket streaming, OAuth and browser-automation fallbacks behind one normalised adapter.

Risk engine

Live P&L over SSE and WebSocket. Automated loss limits, position flattening, account lockout, hedge detection and broker-side rule sync — with explicit, auditable enforcement outcomes.

Automation

TradingView, ATAS and webhook signals validated against per-account risk rules before execution. Master/follower copy trading with cross-broker symbol mapping.

AI in-product

A multi-provider LLM gateway (Anthropic, OpenAI, Gemini, DeepSeek, Grok, Groq) on bring-your-own-key with encrypted storage and trading-context injection. A real-time speech pipeline: PCM audio → Azure speech-to-text → translation → neural TTS, broadcast to concurrent clients.

Production

Docker, Nginx, PgBouncer, Prometheus and exporters, GPG-encrypted backups, segmented networks. Stripe subscription billing with entitlements, trials and reconciliation. GDPR consent. WCAG 2.1 AA targets.

Methodhow the work actually gets done

AI-assisted, but the engineering is mine

I don't write code the way a computer-science graduate does. I led hotel teams for twenty years; now I lead AI coding agents the same way — I direct the work, specify it, review it and own it. That turned out to be a method rather than a compromise. These three steps depend on each other in order; drop the first and the third stops working.

1. 01

   The specification is the source of truth

   Every backend module carries a written behavioural contract — 18 of them. AI implements against the spec; I own the contract, the review and the acceptance criteria. When the code and the spec disagree, the code is wrong.

2. 02

   A toolchain, not ad-hoc prompting

   A private library of 20+ domain-specific agent skills covering broker protocols, risk logic and execution semantics, plus a persistent memory layer, structured session handoffs and a codified lessons file — so context and hard-won fixes survive between sessions instead of being re-derived.

3. 03

   Verification is what makes it safe to move fast

   245 backend test modules, Playwright end-to-end suites, 161 versioned migrations, staged deploys. The system moves real money through real brokers for real customers, so velocity without verification isn't velocity — it's exposure.

Stackwhat I reach for

Languages

Python · TypeScript · JavaScript · SQL · Pine Script · Bash / PowerShell

Backend

FastAPI · SQLAlchemy 2.0 · Alembic · Pydantic · asyncio · REST / WebSocket / Server-Sent Events · OAuth 2.0 · JWT

Frontend

React 18 · Vite · Tailwind · TanStack Query & Table · Zustand · Radix UI · Lightweight Charts · Tauri

Data

PostgreSQL · TimescaleDB · Redis · SQLite · schema design and versioned migrations

Infra

Docker & Compose · Nginx · PgBouncer · Prometheus · Cloudflare · Linux VPS · encrypted backup and restore

AI

Claude Code and agentic workflows · Anthropic & OpenAI APIs · Model Context Protocol servers · prompt and context engineering · Azure Speech

Testing

pytest · Vitest · React Testing Library · Playwright

Domain

Futures and CFD market structure · broker execution protocols · order and position lifecycle · risk and drawdown rules · prop-firm evaluation models

Before this1998 — 2020 · hospitality operations

Twenty years of running things that break in public

This part isn't a detour I'm apologising for. Running a large operation teaches you P&L, hiring, standards, and exactly what it costs when quality slips — with a room full of customers watching. I'd argue it's the reason PropDesk became a product instead of a side project.

• 2020 — Independent futures & CFD trader Paid out by six leading prop firms: FTMO, Topstep, MyFundedFutures, Lucid Trading, Tradeify and Take Profit Trader. Verified FTMO record — 15 payouts, US$39.8K, Max Capital Allocation: public profile. The domain behind every requirement in PropDesk — the product exists because I needed it myself.
• 2016 – 20 Director of F&B · SANA Malhoa Hotel, Lisbon ≈ €2M annual revenue, full P&L and budget ownership. Led recruitment, training and performance standards, and authored the department's SOPs through the unit's restructuring.
• 2012 – 16 Director of F&B · EPIC SANA Luanda Hotel, Angola Five-star, 288 rooms, seven outlets, ≈ US$15M annual revenue. Owned concepts, commercial positioning, budgets, hiring and operational controls for a large multicultural team.
• 1998 – 2012 Premium hospitality & events · Lisbon Restaurante Eleven (Michelin-starred), Casa da Comida, Kais, Grupo Silva Carvalho. Staff and service management for embassies, government institutions and private clients.