Kaden VanHoecke

Dedicated headless server running as a persistent AI assistant with graduated autonomy. Features include: a Prompt Engineering framework with deployable documentation templates, an Active Memory system (brain.db) that maintains session continuity and architectural knowledge across conversations, a curated spaced-repetition learning system with Obsidian integration, an automated 16-category security auditing suite with Sentinel AI analysis, orchestrated AI agents per domain (research analyst, knowledge curator, security sentinel, scribe), a document crawling and embedding pipeline with 5 discovery methods, and a full business management layer (LLC docs, profile, job scanning, platform exports). Accessed via mobile-first PWA over Tailscale mesh.

v0.96.0

CloudBooks is a task tracking and payment management platform built for informal employers — property owners, parents, anyone managing real work without a formal business entity. React 18 + TypeScript frontend with Zustand state, TanStack Query, and Supabase Cloud (PostgreSQL + Auth + Realtime + Edge Functions). Multi-tenant architecture with row-level security, property-scoped views, role-based access control (owner/admin/member/viewer), and a Stripe-powered subscription system (checkout, webhooks, customer portal). Features include invoicing with status workflows, recurring task automation via pg_cron, a command palette, PWA with offline support, and CSV/PDF export. Deployed on Cloudflare Pages with GitHub Actions CI/CD and Lighthouse 90+ scores.

Production multi-tenant SaaS with Stripe billing, 12-table RLS schema, and 278 automated tests

Production automation platform for an HVAC contractor's sales engineering team. Parses manufacturer pricing sheets (Armstrong XML, WaterFurnace/Condair/Marlo PDF) into formatted Excel worksheets, matches 180+ product line templates via a 5-tier pipeline (pattern, pgvector embedding, pg_trgm fuzzy, Claude AI, auto-create), and generates Word scope documents. Syncs 375 projects from SharePoint via Microsoft Graph API, correlates them against a 563K-file R:\\ drive census, and tracks 3,000+ equipment orders from 13 PSR spreadsheets with lifecycle state machine transitions. Built on FastAPI + React/TypeScript/Vite with PostgreSQL (Supabase), served through 7 NSSM Windows services behind a Cloudflare Tunnel. Includes a local LLM analyst (Qwen3 14B via Ollama), filesystem watcher with real-time intelligencealerts, and a 4-agent analysis pipeline for continuous system health auditing.

Reduces hours of manual estimating spreadsheet and document assembly to minutes across a 7-service production architecture

AI design assistant that runs inside SOLIDWORKS as a COM add-in, not a web wrapper or external chatbot. Reads live CAD sessions through 19 typed grounding tools (features, dimensions, mates, properties, rebuild diagnostics, reference graphs), reasons over results using configurable AI providers (OpenAI, Anthropic, or local models), and presents grounded answers in a conversational Task Pane. Write operations follow an 8-step safety lifecycle: plan, preview, approve, apply, verify, trace, undo, and log. Works fully offline with deterministic fallback. Free and open-source. SOLIDWORKS Solution Partner Program application in progress.

Native in-process AI assistant for SOLIDWORKS with governed safety lifecycle

Self-directed market analysis system that generates hypotheses, runs experiments, and adapts its own signal weights through outcome tracking. Combines four independent cortexes (LSTM technical, FinBERT sentiment with discovered anti-correlation inversion, Perplexity LLM fundamental, local indicator consensus) into regime-aware predictions gated by an 8-layer risk system. Runs as a high-frequency daemon (5-min core cycles, 30-min full sweeps, DST-aware scheduling) with paper trading execution via Alpaca, Sentry observability, and a React/TypeScript monitoring dashboard. Built on Python 3.13, FastAPI, SQLite, TensorFlow, and PyTorch. The system discovered and corrected its own biases autonomously: sentiment signals were anti-predictive and got inverted, overconfident predictions were filtered, and SELL signals were quarantined after the system measured them at 33% accuracy.

Autonomous curiosity engine: 4 ML cortexes, 8 risk gates, 2,600+ self-evaluated predictions with adaptive weight learning