Projects
Tools I have built for legal work, local agents and games. Expand a project for details and links.
Games: the Steads, building with my children and what I learn →
Frontier work · active development
From LegalQuants WhatsApp groups to a personal assistant and games administrator in Slack. Clint runs as part of the custom agent work on my home Linux AI server.
The development story covers research, memory, “dreaming”, experiments in self-improvement and what I want this work to make possible.
Two skills released in the LegalQuants plugins at the launch of OpenAI's Astra for Law: testing legal arguments against documents and drafting time entries from work records.
pressuretest sets out a legal position, develops objections and tests them against the supplied documents. It considers whether another argument supports the same conclusion, or whether a supposed inconsistency disappears when the next clause is read. Its report links findings to source material and records objections the documents answer.
timenarratives groups selected work records and the lawyer's account by activity, then drafts time entries for review. It is designed to distinguish the lawyer's work from other people's contributions in an email chain and from text generated by AI.
Both use code for calculations, source-reference checks and output requirements alongside the model's analysis. The lawyer reviews the result.
Systems that check citations against sources and return three verdicts: verified, wrong or not verifiable.
These systems check citations against Find Case Law and BAILII for England and Wales, and CourtListener for the United States. Code extracts citations for checking and records the basis of each verdict. A citation that cannot be verified is reported separately from a confirmed error.
The same verifiers run as a standalone UK auditor, inside a multi-agent checker for US briefs, and inside CaseKit, a local-first desktop application for civil disputes in England and Wales with source-locked drafting and bundle export.
Around twenty applications for group claims, litigation funding and checking AI-assisted work. Calculations run in code.
Settlement waterfalls, damages simulation, Part 36 strategy, disclosure defensibility, witness-statement pressure testing, funder pack generation and costs analysis.
Calculations are tested against known answers. Automated checks cover source attribution and the handling of legal authorities.
A platform for a private community of 150+ lawyers across 17 jurisdictions, with chat, shared resources and a three-dimensional meeting space.
LQuorum provides end-to-end encrypted chat, a process for submitting and reviewing shared resources, connections for members' own agents, and isolated sessions for demonstrating third-party tools. The Spire is a virtual meeting space with spatial voice and AI assistants for finding resources.
Open-weight models running on hardware I own and operate, with document retrieval and monitoring.
A machine with 128 GB of unified memory serves several live products. I manage model versions and memory use, connect the models to source documents and monitor the services through an administration dashboard I built.
I have developed five generations of legal retrieval systems, starting on an eight-gigabyte laptop GPU. Later versions combine keyword and semantic search, rerank the results and divide case law and corporate filings into passages suited to each document type.