Projects

Ideas tested in working code.

Working implementations that bring ideas into practice. Each project explores technical concepts, architectural decisions, and engineering trade-offs that support the case studies. One of them is a product I am building for real use.

Oxaron

A product I am building myself, live at oxaron.com and in active development ahead of general availability. Most organisations cannot answer a simple question about their own software: what are we actually paying for, who is actually using it, and what would one more tool really cost us? Oxaron connects read-only to the identity system staff already sign in with, Microsoft 365 or Google Workspace, and turns that into one view of the whole estate, priced.

What it does: an audit that puts a figure on unused accounts, wasted licences, and overlapping tools; a stack map that shows every tool grouped by the job it does, sized by how many people it covers and shaded by what it costs a year; and a simulation that tests a tool you are considering against everything you already run, before anyone signs for it.

Concepts explored: read-only integration with enterprise identity providers, multi-tenant isolation enforced at the database level and covered by tests, deriving cost signals from directory and usage metadata rather than from content, and treating every figure as something the customer must be able to check, with the rule and the arithmetic shown alongside it.

Why these decisions: the product asks a finance or IT lead to act on a number, which means the number has to survive being questioned. Read-only access with a names-and-numbers scope, never files or messages, keeps the blast radius small and the security review short. Showing the working behind each euro turns a disputed figure into something that can be checked rather than argued about. EU hosting is a procurement requirement in the market it serves, not a detail to retrofit later. And deciding where intelligence genuinely improves a decision, rather than where it would simply be expected, is part of the engineering.

  • TypeScript
  • Multi-tenant SaaS
  • Microsoft 365 & Google Workspace
  • Read-only integrations
  • EU-hosted
01

Document Copilot

A retrieval-augmented assistant that answers plain-English questions over a corpus of SEC filings with sourced, citable answers. Built to test the core idea from the Document Copilot case study: that trust, not raw model intelligence, is what makes AI usable in high-stakes knowledge work.

Concepts explored: hybrid retrieval (vector + full-text search), chunking strategy for long-form documents, grounding prompts so the model only answers from retrieved evidence, and citation as a first-class output requirement.

Why these decisions: retrieval over fine-tuning keeps answers current as filings are added and traceable to source documents; a single Postgres store with pgvector reduces operational burden for a team without dedicated infrastructure; hybrid search catches both semantic meaning and exact financial terms.

  • FastAPI
  • React + TypeScript
  • Postgres + pgvector
  • OpenAI
  • Railway
02

Planned experiments

Implementations on the workbench next:

  • Evaluation framework for groundedness
  • Authentication prototype
  • Prompt optimization experiments
  • MCP integration