Projects

The starting point was other people’s good work. Ivo Szapar showed how a structured repo (CLAUDE.md, INDEX.md, dev-docs) turns a codebase into something an AI agent can actually reason about. Dave Killeen and the team behind Dex demonstrated what a personal CRM looks like when it is built for real relationships, not sales pipelines. I learned from both, then built my own version.

Sune is a local-first, AI-augmented personal knowledge management system. Plain Markdown files. Git for version control. No proprietary platform. No vendor lock-in. Everything exportable in five minutes.

The architecture has five layers. Storage: Markdown in Git. Routing: the PARA method (Projects, Areas, Resources, Archive) decides where each note lives. Navigation: hub pages that connect related notes without forcing a rigid hierarchy. Workflows: daily and weekly reviews that keep the system from going stale. Agent layer: ten AI agents, each with a defined output contract, handling summarization, extraction, and classification.

The key design decision is the staging pattern. AI agents never write directly into the knowledge base. They write to a review zone. A human approves or rejects every piece before it enters the system. This keeps AI useful without making it unsupervised.

Ivo and Dave deserve credit for proving that structured personal knowledge systems work. Sune is my implementation of that idea: plain text, version control, and agents that know their boundaries.

This is a company knowledge base built for a client using a repository-based AI approach. The method structures a repository with clear governance (CLAUDE.md defines rules), navigation (INDEX.md maps the terrain), and living documentation (context, plan, tasks, and decisions in dev-docs).

The setup is not complicated. Product catalogs, pricing rules, customer data, and document templates live in structured Markdown files. AI agents read these files and generate outputs (quotes, summaries, reports) following defined workflows called skills and commands.

What makes it work is the organization, not the AI. The same AI models that produce garbage with unstructured input produce accurate, useful output when the knowledge base is well-organized. The governance file tells the system what it can and cannot do. The index tells it where to find things. The dev-docs give it context about recent decisions and current priorities.

The client went from manual quote creation (1–2 hours per quote, frequent errors in product specs and pricing) to automated generation with human review. Now in 15 minutes. The system handles the routine work. Humans handle the exceptions.

The lesson is transferable to any organization sitting on institutional knowledge trapped in people’s heads, shared drives, or email threads. Structure the knowledge first. The automation follows naturally.

ANN 2.0 takes the original codebases and ports them to modern tools. MATLAB replaced with Python 3.11+, XGBoost, SHAP for interpretability, and Optuna for hyperparameter optimization. Using an AI-assisted workflow, I built a prototype that outperformed the original model—built by four data science specialists—by 26%.

Phase 1 results: Mean Absolute Error dropped 26% compared to the original models. The MASE (Mean Absolute Scaled Error) landed at 0.79, below the ship threshold of 0.90 (where anything under 1.0 outperforms naive forecasting).

Phase 2 is waiting on a data export: 1.4 million production events, 261,000 bookings, across 27 sites. The original theses worked with limited academic datasets. The production data will test whether these models hold up at scale and across locations.

The angle for operations teams: if you have old analytical work gathering dust—theses, pilot projects, proof-of-concepts from five years ago—the tools to make them production-ready are dramatically better and cheaper now. The hard part was always the domain expertise. That part does not expire.

JobSearch is a local-first job search system for senior professionals in the Nordics. It has four subsystems: job ingestion (connectors for JobTech API, Greenhouse, and Teamtailor), AI-assisted CV generation, a lightweight CRM with full audit trail, and a macOS desktop UI built with Electron and React.

The architecture is deterministic by design. Every CV statement traces back to a source document. Every application logs which model generated it, which experience entries it pulled from, and what the human changed before sending. The stack runs on Python 3.11+ and Node.js, with Claude, OpenAI, and Gemini SDKs handling different generation tasks.

The philosophy behind it: five precise applications beat fifty generic ones. The system enforces this by making it easy to write targeted applications and hard to spray-and-pray. The CRM tracks every interaction, so nothing falls through the cracks.

This started with a simple observation. After years of consulting, I know how many hours go into creating tailored consultant profiles and CVs for every new gig. I kept seeing people on LinkedIn describe the same frustration—spending weeks on applications, rewriting the same information, getting silence in return. Job searching is painful for most people. I wanted to see if it could be less so. I wrote about the underlying problem in Tinder for Jobs.