Investors talk to the companies they own all the time, about emissions, pay, board seats. The record of those conversations is scattered across stewardship reports, filings and news, and almost impossible to compare. I built a platform that reads those documents and files every engagement the same way: who started it, how hard they pushed, what it was about, what they asked for, and whether anything checkable came of it.

It has an ingestion service, an operations console and a client portal. The first market was going to be European oil and gas, for investor relations teams, asset managers and regulators.

The same problem as Engagement Intel, turned around. Instead of reading an investor's report on whom they engaged, you start from the company: which investors have been at it, about what, how hard, and whether anything came of it. One page per company, with every engagement aimed at it laid out on a timeline.

It reuses the five-way classification from Engagement Intel, so each engagement arrives already sorted by who started it, how intense it was, the topic, the ask and the verifiable outcome. For an investor relations team it answers "who is coming for us and why"; for an asset manager, "who else is pushing on this".

I wrote down how I want to vote the shares I hold: strong on climate, sceptical of oversized pay, generally supportive of shareholder proposals. Codex reads each meeting notice and proxy statement against that and recommends a vote on every item, with its reasons.

For one 2026 AGM it spotted an accounting restatement and a weakness in internal controls, and recommended voting against the audit committee nominees, the auditor and the pay package.

An agent that trades spread bets through the Capital.com API. It watches live prices across the whole investable universe it's allowed to touch, reads a bank of news and macro APIs, and a language model says whether a setup is worth taking. It never places the trade itself: a separate piece of ordinary code sets the size and the limits and decides whether anything is executed at all. It runs in three modes, shadow, demo and live, with a dashboard that mirrors what it's doing.

Everything it sees and does goes into a large persistent memory. Before each session it reads its own history back and reflects on what worked, so the strategy improves from evidence rather than from me fiddling with it. Next to it sits a Trading 212 connection, so I can review my own portfolio from an AI assistant.

A 90-second audio briefing for compliance teams, out every morning at 07:00 CET: what the European regulators published overnight (ESMA, the FCA, BaFin, FIN-FSA, the AFM), in 13 languages. Each episode gets a severity score from Calm to Acute, which sets the running order and even the music under the voice.

LLM routers promise to send your request to the best model for the price. Nobody checks. Routerouter would: did the router actually use the model it claimed, what did you really pay per successful answer, and how does it behave when a provider has an outage. Independent by design: retail access only, no vendor money, a public methodology and an open-source test harness.

The landing page is built and you can click through it. The numbers on it are placeholders and say so; no router has been measured yet.

Describe someone, and get a one-of-one greeting card: the picture and the words made by AI, printed and posted by hand, for £6.99 with the stamp included. The waitlist is live; the ordering flow and the Cloudflare back end are built.

The part I'd show anyone: before choosing an image provider I ran a blind bake-off. Forty briefs times four creative directions, so 160 test cases. The reviewer couldn't see which provider made which image. One case in the set is a request that should be refused, to check that it is. And nothing costs money until I type a confirm flag, so a bug can't run up a bill.