Law AI
An open legal-intelligence platform for prospective law students, current law students and practicing lawyers: LSAT preparation, guided legal study, and legal analysis grounded in a citation knowledge graph — with provenance on every answer and a human in the loop where it matters.
Who it is for
| Audience | What Law AI gives them |
|---|
| Prospective law students | The LSAT lab: logical reasoning, reading comprehension, logic games and argumentative writing, with solving aids and playbooks, a continuous curriculum, question diagnosis and a knowledge library. Also published as a zero-server static site (bamr87.github.io/law-ai, built from lsat-static/). |
| Current law students | The legal-journey platform: matters, scanned-page intake from a book (the ocrmd plug-in), an analysis studio, and agent traces that show how an argument is built, step by step. |
| Practicing lawyers | IRAC analysis grounded in retrieved authority, precedent hunting, a Neo4j citation knowledge graph (GraphRAG: lexical retrieval plus CITES traversal), an audit trail for every model call and retrieval, human-approval gates on ingestion and high-stakes analysis, and a fully local mode (LLM_PROVIDER_CHAIN=ollama) for privileged data. |
Every output names the authority it rests on, labels asymmetry rather than filtering by it, and refuses honestly when it cannot ground an answer (docs/philosophy.md, docs/constitution.md §13).
How it works
- Backend — Django 5 (ASGI via daphne), DRF, Channels (WebSocket agent streaming), Celery/Redis (async ingestion), Neo4j (knowledge graph) + Postgres + Redis. Each Django app owns one
/api/* router: agentsapi, analysis, graph, lsat, matters, ingestionapp.
- Agents — CrewAI crews and LangGraph state graphs (
backend/law_ai/agents_runtime/), running on ChatProviderChain, a LangChain BaseChatModel over the product's LLM provider chain (backend/law_ai/core/llm/): Anthropic Claude (subscription OAuth token, then ANTHROPIC_API_KEY), then xAI Grok (subscription OAuth, make xai-login), then local Ollama as the last tier. Pin LLM_PROVIDER_CHAIN=ollama for local-only operation — required for any data under a confidentiality obligation that forbids third-party processing (docs/constitution.md §1, docs/llm-provider-chain.md).
- Frontend — Next.js 16 App Router, TypeScript strict, Tailwind v4, Cytoscape.js. Two UI apps (
frontend/, lsat-static/) on a first-party design system (docs/ui-ux.md).
- Infra — Docker Compose locally (Postgres, Neo4j, Redis, Ollama, Phoenix tracing), GitHub Actions in CI.
Full architecture: docs/architecture.md · agents: docs/agent-roles.md · data model: docs/data-model.md · product requirements: docs/prd.md.
Quick start (5 minutes to running)
git clone https://github.com/bamr87/law-ai.git
cd law-ai
cp .env.example .env
make dev # or: docker compose up --build
# frontend: http://localhost:3000 API: http://localhost:8000 (routes under /api/agents, /api/analysis, /api/graph, /api/lsat)
Without Claude credentials in .env the provider chain falls through to the local Ollama tier automatically — you still get answers, from the local model. Add CLAUDE_CODE_OAUTH_TOKEN (from claude setup-token) or ANTHROPIC_API_KEY to use Claude.
Everything runs in Docker. Every dev/test/lint command is make … or docker compose …; there is no supported host-native toolchain. To wire your editor's language server to the same environment CI uses, open the repo in one of the Dev Containers (Law AI (backend): .devcontainer/devcontainer.json; Law AI (frontend): .devcontainer/frontend/devcontainer.json).
make test # backend pytest + frontend vitest
make lint # ruff + black + eslint
make typecheck # mypy --strict (ratcheted) + tsc
make lint-specs # the spec/backlog gate (the fleet's, via tools/fleet-kit.sh) + the UI shell audit
make test-legalbench
Project structure
| Path | What it is |
|---|
backend/ | Django (ASGI) API, agents, the LLM provider chain, ingestion, the LSAT content factory |
frontend/ | Next.js App Router platform app |
lsat-static/ | Static LSAT study site published to GitHub Pages (make lsat-export snapshots the bank) |
lawmode/ | The imported bamr87/lawmode legal reviewer for developers (its own README and tooling; runs from inside lawmode/) |
specs/, BACKLOG.md | The spec packages and the append-only backlog — the record of what shipped and what is next |
docs/ | Architecture, data model, agent roles, the constitution, ADRs, the PRD, runbooks; docs/history/ keeps the original proposal and planning notes |
infra/, docker-compose*.yml | Local stack: Postgres, Redis, Neo4j, Ollama and observability |
scripts/, tools/, tests/ | Repo-level stdlib scripts (UI audit, visual-evidence gate, mypy ratchet, LegalBench and Ollama helpers), hub-vendored tooling (fleet-kit.sh, unwrap-prose.py), and the repo-wide structural tests |
Working on Law AI
AGENTS.md is the canonical guide for humans and AI agents alike: the locked stack, the layout, the Docker-first rule, the coding and testing standards, and the gates every PR must pass. GitHub Copilot and Claude Code are the sanctioned AI development surfaces; other coding agents (Cursor, Continue.dev, Aider, Codex, Gemini CLI) are not supported and their config files are rejected at review.
Development follows the fleet's spec-driven loop rather than a workflow of this repo's own: a non-trivial change starts as a spec package under specs/ (/spec-driven:specify → :clarify → :plan → :tasks → :implement, the spec-driven Claude Code plugin from the bamr87 hub), the spec and backlog gates run by reference in CI (Spec Drift Gate) and locally as make lint-specs, and every cycle mirrors the follow-ups it discovered into BACKLOG.md. docs/spec-driven.md maps the loop onto this repo and lists which checks block a merge.
All contributions, human or AI-generated, pass make lint + make typecheck, make test (backend coverage ≥75%), make lint-specs, the relevant LegalBench tasks (required on PRs labelled reasoning-change, weekly otherwise), screenshots for anything a user could see, and a review of the agent traces for auditability.
Project status
Law AI is under active spec-driven development — there is no fixed roadmap snapshot here, because it drifts. For what is shipped, in flight, and planned:
- Shipped features: the numbered spec packages under
specs/ (each spec.md + plan.md + tasks.md), the source of truth for what exists.
- What's next: the
## Open section of BACKLOG.md.
- LegalBench regression gate:
backend/law_ai/eval/legalbench.py, run via make test-legalbench / make legalbench-baseline.
- Changes: CHANGELOG.md and the commit history.
License
MIT.
Part of the bamr87 portfolio — browse every project on Amr's GitHub profile.
