Rich Berman
Rich Berman
Legal technology consultant and systems builder focused on reliable automation for law firms.
I build the layer between legal operations and software: intake systems, workflow automation, integrations, reporting, document automation, AI-enabled tools, and the controls that keep those systems inspectable and useful in production.
My public work emphasizes a simple idea: use models for language and judgment support, and use deterministic software for rules, records, gates, provenance, and verification.
Engineering methodology: how I use AI agents to build deterministic systems without trusting the agents to be deterministic.
Selected work
| Project | What it demonstrates |
|---|---|
| intake-triage-mcp | A deterministic MCP server for legal intake triage with structured validation, provenance, a hard conflicts gate, adversarial regression tests, and append-only logging. No model or network calls in the decision path. System proof |
| intake-eval-harness | A reusable MCP evaluation harness for answer and execution-trace assertions, JSON/JUnit evidence, provenance, latency and call budgets, and regression gates across stdio, SSE, and streamable HTTP. |
| mhsb-intake-leak-calculator | A browser-only law-firm intake model with explicit assumptions, sourced coefficients, automated tests, accessibility checks, and zero runtime tracking. Live tool |
| llm-security-for-law-firms | A practical threat model and adoption checklist for using LLMs in law firms. Read it |
| granolacowboy.dev | Source for my technical field notes and case studies, built as a minimal static Astro site with build-time verification. Visit |
Engineering principles
- Deterministic where failure matters. Conflicts gates, validation, calculations, audit records, and policy enforcement should not depend on a model behaving itself.
- Evaluate systems, not demos. Golden cases, repeatable tests, CI, smoke checks, and explicit failure semantics matter more than impressive one-off outputs.
- Preserve provenance. A useful answer should make it possible to identify what data, rule, assumption, or source produced it.
- Keep humans in the control plane. Automation should surface decisions and exceptions clearly instead of hiding them behind a label like "AI".
- Minimize unnecessary data movement. Local-first and browser-only designs are preferable when the workflow does not require a remote service.
- Measure operational outcomes. Technology should improve throughput, quality, response time, consistency, or decision visibility, not merely add another interface.
Current work
- MHSB Solutions: legal-technology strategy, implementation, workflow automation, integrations, reporting, and AI enablement for law firms.
- LexLabs: productized Lawmatics implementation systems.
- efficient.esq: law-firm AI operating-model and governance work.
- Defensive security research: hardening the systems, agents, and infrastructure used to run the above safely.
Public proof chain
The flagship intake work is deliberately split into inspectable layers: system demonstration, then the deterministic MCP, then the golden suite, then the evaluation harness, then release evidence. Start at the system demonstration and follow the links through each layer.
Research library
stars is my automatically maintained GitHub research index: thousands of repositories organized into topic-specific lists across AI, agents, security, automation, infrastructure, legal technology, and adjacent tooling.
Writing
- How I use AI agents to build deterministic systems without trusting the agents to be deterministic
- Anatomy of a legal intake automation
Contact
- Website: granolacowboy.dev
- GitHub: github.com/granolacowboy
- Hugging Face: huggingface.co/granolacowboy
- Kaggle: kaggle.com/granolacowboy
- MHSB Solutions: mhsbsolutions.com
- LinkedIn: linkedin.com/in/mhsb
- Email: rich@mhsbsolutions.com