Jonathan Kleiman Email me
Toronto · Called to the Ontario Bar, 2026

I turn legal tasks into systems AI agents can run.

Ontario lawyer and AI builder. I built Canadian legal research for AI agents, an agent harness that nearly doubled all-pass rates on Harvey's Legal Agent Benchmark, and an open model that reads how courts treat the cases they cite.

See the work Résumé (PDF)
01 / Workdelphi2026 – present

Delphi

Private beta

Canadian legal research for AI agents. An MCP server that gives Claude and ChatGPT 1.7 million Canadian cases and statutes with paragraph pinpoints, a 16-million-link citation graph, a built-in citator, and a checker for invented citations.

example output

delphi-mcp.com
02 / Workmithril-harnessMay – Jun 2026

Mithril Harness

An agent harness for Harvey's Legal Agent Benchmark. A criteria-blind loop that must read every source document, with fresh-context reviewers rechecking figures, dates and citations. It brought open-source models up to the level of closed frontier models.

github.com/jmklex/mithril-harness
03 / WorklevOct 2026

Lev

Apache-2.0

An open 9B decision model for case law. Fine-tuned d3-flash (Qwen3.5-9B backbone) to classify Canadian court decisions for citator work, such as how a court treats a case it cites.

huggingface.co/jonathankleiman/lev
04 / Hardwareasus-googlebook-omarchyOct 2026

Linux on a Googlebook

MIT

The first native Linux install on the ASUS Googlebook 14. A full Omarchy desktop boots from the encrypted internal drive on the factory firmware, with no replacement firmware and no UEFI conversion. I turned it into a reproducible image builder with staged install and recovery tools.

DisplayKeyboard + touchpadWi-FiGPUSpeakers

verified on hardware · 7 Oct 2026

github.com/jmklex/asus-googlebook-omarchy
05 / Methodlegal_task → agent_workflow4 steps

How I turn legal work into agent work

  1. 01 decompose

    Break a legal task into decisions, rules, exceptions, and the points where a person has to decide.

    BLG: turned recurring litigation drafting into repeatable automated workflows.

  2. 02 define_good

    Write the examples, criteria and test cases an agent is judged against.

    Mithril Harness: generation stays blind to the criteria; separate reviewers check the output against the sources.

  3. 03 find_failures

    Run realistic scenarios, find where the agent breaks, and fix the cause.

    Delphi: a checker that catches invented citations before anyone relies on them.

  4. 04 prototype

    Get from idea to a working prototype, then hand engineers something real.

    Lev: from training run to released weights, code and report.

Tools

  • Claude Code
  • Codex
  • MCP servers
  • RAG
  • LLM evaluation
  • LoRA fine-tuning
  • PyTorch
  • vLLM
  • Python
  • JavaScript
  • Ruby
06 / Experience§ law2023 – present

Where I've worked

  1. Borden Ladner Gervais LLP

    Articling Student, Summer Law Student

    Summer 2024,
    2025 – 2026

    Researched and wrote on AI governance and compliance (Canada's AIDA, the EU AI Act). Drafted pleadings, memoranda and court orders in civil procedure, class actions and health law, and built repeatable automated workflows for that drafting. Argued motions and appeared before arbitration panels.

  2. ZeroDraft AI (simplegrants.ai)

    Junior Developer & Compliance Officer, Co-founder

    Sept 2024 – present

    Compliance procedures for an AI startup in the regulated grants and tax-credit space. LLM workflows behind production grant-matching and grant-writing features. Conferences, procurement and go-to-market.

  3. Community Legal Education Ontario

    Summer Law Student

    Summer 2023

    Access-to-justice research; briefing notes, reports and policy papers.

University of Toronto, Faculty of Law · JD, after 1L at Osgoode Hall2025

University of Toronto · History, Dean's List2018 – 2022