I turn business processes into working AI systems.

Riskeys

Trading-analytics SaaS

Founder & lead engineer

A Laravel platform that turns raw MetaTrader trade history into honest analytics. Traders run a custom client that streams their account state in real time; the backend computes performance, normalizes broker time, and surfaces behavioral patterns.

• Custom analytics engine: six metric computers with partial-close grouping so a multi-leg exit counts as one trade.
• Broker-clock normalization detects each broker's UTC offset and re-bases every timestamp before time-bucketing.
• Seven behavioral pattern detectors (overtrading, revenge trading, discipline drift) on a clean interface/DTO contract.
• Real-time trade sync: the client pushes account state as JSON every 100ms; the server reconciles open and closed positions.

87

models

193

migrations

98

tables

101

controllers

• Laravel 10
• PHP 8.1
• MySQL
• Stripe
• Shopify
• Telegram
• Twelve Data
• Resend

Riskeys EAs

Native MetaTrader software

Sole developer

The desktop side of Riskeys: trading-risk software written natively in MQL4 and MQL5. An on-chart panel for position sizing and journaling, talking back to the platform over a hand-built HTTP layer. Not an auto-trader; the trader always decides.

• Custom on-chart GUI framework from scratch: 30+ classes for windows, panels, buttons and event handling.
• WinAPI HTTP layer plus a JSON parser in MQL for token auth, trade upload and screenshot POST.
• Prop-firm rule engine validating position sizes and daily limits from per-broker JSON configs.
• MT4 and MT5 kept in parity through a documented dual-repo git workflow.

~78K

lines MQL

50+

modules each

30+

GUI classes

2

platforms

• MQL4
• MQL5
• WinAPI
• JSON
• Inno Setup

ContentOS

AI content platform

Architect & builder

A Next.js + Supabase system that generates on-brand marketing content grounded in real product data. A dual-LLM design keeps publishable copy in the cloud and source-code analysis fully local, with an automated checker that blocks false claims before anything ships.

• Dual-LLM split: Anthropic Claude for publishable copy, local Ollama for code analysis that never leaves the machine.
• RAG pipeline ingests a codebase, chunks and embeds it with pgvector, and extracts marketing-safe product claims.
• Automated claim and compliance checker validates every generated piece against a policy database.
• MCP-based agent fulfillment: the app queues work, an agent pushes it (e.g. to Klaviyo).

103

tables

~3K

lines SQL

768

dim embeddings

2

LLM tiers

• Next.js 14
• TypeScript
• Supabase
• pgvector
• Anthropic Claude
• Ollama
• MCP

Hermes Graph Memory

Persistent memory layer for an AI agent

Sole developer

A graph-based memory system for a personal AI agent: entities, relations and aliases persisted in SQLite, recalled automatically into every prompt via a submit-time hook. Built with an actual test suite, not just scripts.

• Entity/relation/alias graph schema in SQLite, with a recall hook that injects relevant memory into every agent prompt.
• Layered architecture (persistence, interfaces, backend) with an explicit layering check script to keep boundaries clean.
• Migration system for evolving the schema, plus a corpus ingestion pipeline for backfilling historical context.
• 7 test modules covering the doctrine service, graph runtime, memory service, migrations, tasks and tool registry.

50

commits

7

test modules

~2.5K

lines Python

1

recall hook

• Python
• SQLite
• Graph modeling
• pytest

NBHQ Company OS

LLM-agnostic operating layer for the business

Architect & builder

The organisational layer that lets any AI agent understand and operate across No Brainer HQ's ventures: department charters, registries, and an operating model expressed as structured, agent-readable documents instead of tribal knowledge.

• 119 structured markdown documents covering 6 ventures, each with its own department charters and operating docs.
• Registry system that gives any LLM a canonical, up-to-date map of infrastructure, ownership and process, no re-discovery needed.
• Deliberately LLM-agnostic: works the same whether the operator is Claude, GPT or a local model.
• Ops runbooks per host (hermes-host, nbhq-server, riskeys-vps) so infrastructure knowledge survives beyond any one session.

119

docs

6

ventures covered

3

ops runbooks

1

canonical registry

• Markdown-as-data
• Multi-agent orchestration
• Documentation architecture

Tradelingo

Duolingo-style trading education platform

Architect & builder

A behavioral trading-education platform built from Riskeys' 140-day curriculum, turning a static course into an interactive, gamified learning product with its own product-design pipeline from research to shipped MVP.

• Full product pipeline documented end to end: market research, UX research, IA, personas, journeys, wireframes, hi-fi design and handoff.
• Next.js + Supabase app with Clerk auth and internationalization (next-intl) built in from the start.
• 11 Supabase migrations defining the learning-progress and content data model.
• ~13K lines of TypeScript across a real apps/ workspace, not a single-page prototype.

140

day curriculum

~13K

lines TypeScript

11

DB migrations

21

documented design phases

• Next.js
• Supabase
• Clerk
• next-intl
• TypeScript