I ship AI to production in finance.

An investment research and advisory company with one human and a team of AI agents. Free educational articles and weekly market insights, and a personal AI advisor on Telegram that turns your goals into an ETF portfolio.

• Nine Claude Code subagents in five departments, coordinated by written workflows instead of AI managers
• Agents hand work over through files in the repository, so every brief, draft and objection is versioned
• Two human gates per article: what to write about, and whether it ships
• An advisor built on goals-based wealth management, with the maths done by code, not the model

An AI investment analyst that runs on your own machine. Ask whether a company is worth owning and it does the work: pulls the financials, runs a written valuation procedure and checks the answer against what you already hold.

• Fifteen written analysis procedures, from earnings quality to margin of safety, versioned in the open
• Eight free public data sources: no paid vendor and no API keys
• Remembers your goals, runs scheduled reviews, and can connect to Alpaca, Coinbase or Interactive Brokers, read-only if you prefer
• Use it from Claude Code, Claude Desktop, any MCP client, a local web app or its REST API

Agent conversations are per session and disposable. Loci keeps what they produce and files it by the work it belongs to, so the next session, or the next person, starts from what was already settled.

• Captures every Claude Code session in a repository through hooks
• Agents save decisions and their reasons as memories, attached to a project, a task and tags
• Search by text or by meaning, with embeddings from a local model
• One static binary, one database file per store, nothing leaves your machine

A pipeline of agents that takes a ticket from a rough brief to a monitored production change. The engineer decides at two points: what to build, and whether it ships.

• A spec review runs before any code exists, so a broken spec costs minutes instead of a full run
• Every loop has a cap, and the run hands back to the engineer instead of guessing
• The spec's production signals become live checks after deploy