Multi-tenant task tracking and business management SaaS for informal employers: property owners, parents, and anyone running real work without a formal business entity. React 18 + TypeScript + Tailwind on Vite 6, backed by Supabase Cloud with RLS on 25 tables, Realtime subscriptions, and 9 Edge Functions. 21 domain service modules, pgcrypto-encrypted AI keys and OAuth tokens, Stripe subscription billing gated at the database via the check_feature_limit RPC, and QuickBooks Online OAuth2 with bidirectional sync. 288 unit and 61 Playwright E2E tests, GitHub Actions CI/CD, PWA installable, deployed to Cloudflare Pages.
React 18 TypeScript Tailwind CSS Vite Zustand TanStack Query React Router v6 Supabase Auth
McQueenyML
Frontend: React built Parses manufacturer pricing sheets (Armstrong XML, WaterFurnace, Condair, Marlo, Johnson Controls PDFs) into formatted Excel workbooks, matches line items against 195 product scope templates through a deterministic-first pipeline (pattern, fuzzy, pgvector embedding, then AI fallback), and generates MGI or JCI Word scope documents. React 18 + TypeScript frontend, Python 3.11 + FastAPI backend, PostgreSQL via Supabase with pg_trgm and pgvector, and Ollama Qwen3 14B for project lifecycle analysis. A real-time R:\ drive filesystem watcher correlates events to 419 SharePoint-synced projects, 3063 PSR equipment orders, and a 394K-row network drive file index. 7 NSSM Windows services run the stack in production, fronted by Cloudflare Tunnel with LAN-routing bypass for uploads over 100MB. ~111 API endpoints across 21 routers, 223 pytest and 55 vitest tests.
Frontend: React TypeScript Vite Tailwind CSS. Backend: Python 3.11 FastAPI uvicorn. Database: PostgreSQL via Supabase (pg_trgm pgvector). ML: Ollama (qwen3:14b nomic-embed-text). Infrastructure: 7 NSSM services on Windows 11
Adze: AI for SOLIDWORKS
C# .NET 4.8 built Adze is an in-process SOLIDWORKS COM add-in that reads the active CAD document through 18 typed tools (feature trees, dimensions, mates, custom properties, rebuild diagnostics, reference graphs, closed-file search) and reasons over the results with an agentic tool-calling loop. Write operations follow an 8-step safety lifecycle: plan, preview, approve, apply, verify, trace, undo-label, history. Runs against OpenAI, Anthropic, OpenRouter, Ollama, or LM Studio, with a deterministic fallback when no provider is configured. Built on C# .NET 4.8 and the SOLIDWORKS API. 666 NUnit tests across broker orchestration, tool implementations, write infrastructure, and closed-file OLE indexing.
C# .NET 4.8 SOLIDWORKS COM API NUnit 3 MSBuild PowerShell WinForms WebBrowser
Oracle (Trade Signal)
Python 3.13 (miniconda3) built Self-directed market analysis system that generates testable hypotheses, backtests them against historical data, and manages a $499K paper portfolio behind 8 risk gates. Five cortexes (LSTM, FinBERT sentiment, Perplexity LLM deep analysis, rules-based momentum, and a local indicator ensemble) feed an adaptive-weight Oracle brain that reweights across 12 market regimes. The distinctive finding, surfaced by 2,987 logged predictions with outcome tracking, is that the system's own BUY signals are anti-predictive for going long at 27% accuracy but produce a 62.8% win rate as contrarian shorts with 2:1 W/L ratio. A multi-venue adapter layer (Alpaca equities wired live, plus verified fee models for Coinbase, Polymarket, and Kalshi) routes signals to whichever venue fits. Python 3.13, FastAPI, SQLite WAL, systemd daemons, Sentry observability, 506 tests across 22 files.
Python 3.13 (miniconda3) FastAPI SQLite TensorFlow PyTorch transformers (FinBERT) React/TypeScript/Tailwind (ava_hub UI)
