Three years finding what breaks. Now building what doesn't.

Problem

One model rarely gives advice you can verify on implementation, security, design and quality.

Approach

A local runtime detects the available providers, chooses which models to involve and learns from results. Statuses separate planned work from executed work.

Stack · status

• Swift
• SwiftUI
• Node.js
• LLM orchestration

Problem

Provider results arrive in different shapes, so comparing them by hand is slow and inconsistent.

Approach

A Turborepo monorepo: Next.js front end, NestJS API and Temporal workflows over normalised provider results. The first demo runs on declared mock fixtures.

Stack · status

• Next.js
• NestJS
• Temporal
• PostgreSQL
• Docker

Problem

A single requirement can take 30 to 60 minutes to cover, and quality varies between people.

Approach

Upload a PDF, DOCX or text requirement and get happy paths, edge cases, negative flows and boundaries, ready to export. It comes from the work I did by hand for three years.

Stack · status

• Python
• React
• Docker
• LLM

Problem

AI-written content needs review, originality checks and a person who signs off.

Approach

A LangGraph supervisor, multi-agent review with plagiarism checks and mandatory human approval. n8n handles publishing, Postgres with pgvector holds memory.

Stack · status

• LangGraph
• Python
• n8n
• Postgres

Problem

Logging repetitive defects by hand is slow, and duplicates slip through.

Approach

Reads defects from Excel, ranks similar existing tickets with embeddings and creates incidents through browser automation. 67 unit tests, no browser needed.

Stack · status

• Python
• OpenAI Embeddings
• Playwright
• pytest