I'm a practicing technology lawyer and legal engineer. I work on the unglamorous half of "legal AI": making a model's output verifiable, auditable, and safe to rely on unsupervised — not just fluent.

The through-line across everything here is deterministic verifiers and honest evaluation. A legal answer you can't check is a liability, so I build the checkers.

• gdpr-dpa-art28-env — a deterministic, verifiable reward environment that scores GDPR Article 28(3) Data Processing Agreements clause-by-clause, with evidence spans and adversarial evasion tests. A worked example of RL-with-verifiable-rewards over legal text.

• counselbridge — grounded legal chat with enforced citations, fail-closed attorney-review routing, and a hash-chained audit trail.
• LeBlox — a privilege-preserving workbench: per-matter isolation, encryption, and DLP that blocks cross-matter leakage.
• plainclause — local-first contract-clause explanations and guarded rewrites, no cloud required.
• legal-prompt-router — a deterministic router mapping a legal-workflow prompt library to a firm's actual negotiating positions.

• leto — a local-first, Telegram-first Python agent framework (typed reasoner, delegation, 56 tools), runnable fully keyless.
• deep-agent-lab — a small, safety-hardened research agent (LangGraph), exposed as both a CLI and an API.
• onboard — a wiki compiler for company knowledge (not a RAG chatbot).
• highland — Mission Control for long-horizon, AI-assisted software builds (Elixir / Phoenix LiveView).
• reactor-lite — AI-driven cyber crisis simulations for security & compliance teams.

• Honest metrics. If a number can't be reproduced from the repo, it doesn't go in the README.
• Deterministic where it counts. Verifier/regex gates over LLM judgment for anything safety-critical — auditable and reproducible by construction.
• Abstention over confident-wrong. The trustworthy primitive is knowing when not to answer.