My personal site: the things I build, and notes on building them. I work mostly with law firms through MHSB Solutions.
A small evaluation harness for tool-using MCP servers: run a fixed suite against a server, score the final answer, and enforce deterministic assertions over the tool trace. Tool-using systems fail in ways a unit test does not catch (the model picks the wrong tool, skips a required step, or answers without calling anything), so a fixed suite turns those failures into a repeatable number instead of a surprise in production. Extracted from intake-triage-mcp.
A standalone, static browser tool: a law firm owner enters their intake numbers in about three minutes and sees an estimate of the annual revenue they could recover by closing intake gaps, broken down across five stages (unanswered calls, slow first response, insufficient follow-up, consult no-shows, and unsigned engagement letters). Zero backend, no analytics, no tracking, no email capture; every computation runs in the browser. The framing is deliberately conservative, with one sourced coefficient (the Clio 2024 answer-rate figure) and every other default badged as an assumption and overridable.
A curated snapshot of my GitHub Lists: 2635 unique repositories across 28 public lists, rendered into a browsable research library. Selection is manual (I decide what belongs in each list); the rendering, the ordering by star count, and the metadata are mechanical and point-in-time. It is a working library, not a ranking or a recommendation engine, and inclusion is not endorsement: the lists hold tools I am evaluating, comparing, monitoring, or may want to revisit.
A concrete teardown of a legal intake workflow, its deterministic tools, human gates, and evaluation strategy.
A reference implementation of secure legal intake automation on the Model Context Protocol. The server is fully deterministic (no LLM calls, no network calls) and handles conflict screening, matter validation, follow-up templating, and triage logging. The language model does the language work, the server does the record work, with a clean boundary between them.
A reproducible walk-through showing why a correct final answer is not enough: conflicts provenance, write refusal, adversarial inputs, trace assertions, and release evidence all sit outside the model.
A public, client-scrubbed research artifact built from aggregate patterns across 1,299 agent sessions. It tests whether an agent can preserve constraints, privacy boundaries, evidence, and executable next steps when the source context is messy.