LexIntel
Live demo: huggingface.co/spaces/Elisha622/lexintel-ai · Source: GitHub · Mirror: Hugging Face
Court case management and legal-intelligence platform for the UAE — decision
support for courts, prosecutors and lawyers. Bilingual (Arabic / English,
full right-to-left), built on the official UAE Design System
(@aegov/design-system), with every AI feature backed by a non-AI fallback.
Humans decide. LexIntel never issues verdicts and never scores emotion or
"deception". Only a judge can enter a ruling. See
docs/DESIGN_DECISIONS.md.
What it does (all modules are real, end to end)
| Module | What you can do |
|---|---|
| 1 · Complaints | Citizens file online (no account), get a reference number + tracking code, and track status. Staff triage with AI suggestions (category, department, duplicates, priority), then open a case in one click. |
| 2 · Case intelligence | Extract people, places, dates and amounts from any text; dated events go onto the case timeline. A case brief quotes the key sentences of the file, the topics it discusses and the laws it cites, with an optional checked draft. Compare two documents/statements for factual differences (never credibility). |
| 3 · Evidence | Upload PDFs, scans, photos, text, audio/video. SHA-256 fingerprint on upload, Arabic + English OCR, transcription of recordings, chain-of-custody log, integrity re-check, review/flag. |
| 4 · Prioritisation | Transparent weighted score from factual signals, with a factor-by-factor explanation; re-scored nightly as cases age. |
| 5 · Similar & related cases | Similar wording across case files (Elasticsearch); people who appear in other cases and cases citing the same law (Neo4j graph). |
| 6 · Legal research | Answers only from the Law Library (official PDFs you upload), with article citations and an in-force date filter. Keyword + semantic retrieval. |
| 7 · Scheduling | Hearing calendar (day/week) that blocks courtroom and judge double-booking. |
| Courtroom stand | One person at a time: clerk confirms identity, real microphone/camera recording, live local speech-to-text (Arabic/English), transcript review, full-quality re-transcription. |
| 9 · Signatures | Signature presence + similarity check against a reference (document authentication only). |
| 10 · Analytics | Live dashboard and analytics with table views for every chart. |
| 11 · Citizen portal | Public site, complaint filing and tracking. |
| 12 · Explainable AI | Every AI output shows its reasons, sources and confidence. Drafts are checked sentence by sentence against their sources, with unmatched sentences marked. All AI runs on this server. |
| Security | JWT sign-in, 5 roles, lockout after 5 failed attempts, tokens revoked on password change, Emirates IDs stored only as keyed hashes, audit log of every action, admin user management. |
Live demo (Hugging Face Space)
The whole platform runs in a single container on a free Hugging Face Docker
Space, built from the root Dockerfile: nginx serves the
frontend and proxies the API, with Postgres, MongoDB, Redis and the local
writing model (Ollama, qwen2.5 1.5B) running beside it. Elasticsearch, Neo4j
and the Celery worker are off there; the app uses its fallbacks. Startup
lives in deploy/huggingface/start.sh.
Sign in with any demo account below and the password LexIntel@2026
(set the DEMO_PASSWORD Space secret to change it). Everything in the demo
is fictional, is shared by every visitor and is reset whenever the Space
restarts. It is CPU-only, so AI drafts take a while, and it sleeps after 48
hours without visitors (the first visit then wakes it in a minute or two).
Run it
Requirements: Docker Desktop. Everything else runs in containers.
docker compose up -d --build
- App: http://localhost:3005
- API: http://localhost:8005 (docs at
/docs)
First run only — create the demo accounts and demo records:
docker compose exec backend python -m scripts.seed_demo --password "LexIntel@2026"
This creates the demo accounts, five clearly-marked demo cases with documents (read by the same pipeline as uploaded evidence, so briefs, key sentences, topics, cited laws and similar cases have material), three demo complaints, and three short sample law texts so the Law Library and Legal research have something to answer from. The sample texts were written for the demo and say so on their first line -- they are not UAE legislation.
Re-run it with --refresh after an update: that refreshes the demo records,
adds anything a newer version of the seed introduced, and moves the demo
hearings back to today, so the Courtroom stand always has a session to open.
| Username | Role |
|---|---|
admin |
Administrator (users, audit log, system status) |
judge.demo, judge2.demo |
Judge (the only role that can enter a ruling) |
clerk.demo |
Court clerk (complaints, scheduling, courtroom stand) |
officer.demo |
Case officer (cases, evidence, triage) |
prosecutor.demo |
Prosecutor (evidence, research) |
All demo accounts use the password you pass to seed_demo. Change them before any real use.
Legal research needs the Law Library
Official UAE statute portals block automated downloads, and this project does not circumvent that. Download the official PDF yourself (for example from uaelegislation.gov.ae) and upload it under Law library. Articles marked "Article (N)" / "المادة (N)" are indexed individually.
The seed ships three short sample laws so the feature can be tried immediately. They are demo text, not real statute: every title starts with "Demo:" so an article cited in a research answer can't be mistaken for a real one. Delete them once you have uploaded the real texts.
AI runs on this server
No cloud AI and no API key. Small task models handle classification, duplicates, similar cases, summaries and offence mentions; a local writing model (qwen2.5 1.5B via Ollama) drafts research answers and case briefs, streamed as it writes and checked sentence by sentence against the sources. Download the writing model once (about 1 GB):
docker compose exec ollama ollama pull qwen2.5:1.5b-instruct
Without it, everything else still works: research shows the in-force articles
and key passages, and case briefs show the quoted key facts. See
docs/ARCHITECTURE.md → AI on this server.
Services and what happens if one is down
| Service | Used for | If it's down |
|---|---|---|
| PostgreSQL | Cases, complaints, hearings, people, rulings, audit | API returns a clear 503 and retries |
| MongoDB | Courtroom sessions, statements, evidence text | Courtroom pages show "temporarily unavailable"; text falls back to Postgres |
| Redis + Celery worker | Background jobs, nightly re-scoring and re-indexing | Jobs run inside the API process |
| Elasticsearch | Full-text search, duplicates, similar cases, law keyword search | Database search and word-overlap matching |
| Neo4j | Relationship graph | Related cases computed from Postgres |
| faster-whisper | Speech-to-text | Clerk types/corrects the transcript |
| Ollama (writing model) | Research drafts, case briefs | Articles + key passages; quoted key facts |
Live status of all of these: System status (admin).
Memory
The full stack idles at about 2 GB, and the writing model adds ~0.9 GB while
loaded. Each service has a memory cap, and the AI models refuse to start when
free memory is short -- you get the quoted sources instead of a draft, which
is far better than the machine seizing up. System status shows the free
memory and what it currently allows. On an 8 GB machine, stop other Docker
projects while using LexIntel, or give Docker more memory in
%UserProfile%\.wslconfig (for example memory=5GB).
The API does not watch for code changes. After editing backend code:
docker compose restart backend
For hot reload while developing, add the dev override:
docker compose -f docker-compose.yml -f docker-compose.dev.yml up -d backend.
Tests
docker compose exec backend python -m tests.run_all
Uses separate lexintel_test databases, never your data. Includes a security
suite (roles, ruling boundary, lockout, token revocation), a full flow suite
run with every optional service switched off (proves the fallbacks), a suite
for the AI task models (classification, duplicates, offence mentions,
summaries, the grounding check and the model slots), and a live Neo4j test.
The runner unloads the writing model first, because holding it costs ~1.1 GB.
Drafting by the local writing model is slow on CPU, so it runs only when asked:
docker compose exec backend python -m tests.run_all --live-ai
Project layout
backend/
app/api/ auth_routes.py + routers/ (complaints, cases, evidence,
scheduling, courtroom, research, insights)
app/agents/ LangGraph agents: intake triage, case intelligence,
legal research, prioritisation, courtroom session
app/db/ Postgres repositories + migrations, Mongo, Elasticsearch,
Neo4j -- each with a reachability probe
app/ingestion/ OCR + signature checks, Law Library parsing/indexing
app/stt/ local speech-to-text (faster-whisper)
app/tasks.py background jobs (Celery, with in-process fallback)
scripts/ seed_demo.py, seed_admin.py
tests/ run_all.py + suites
frontend/
src/components/ UAE Design System components, layouts, charts
src/pages/public/ landing, file/track complaint, sign in
src/pages/app/ dashboard, cases (+ case workspace), complaints,
hearings, courtroom stand, research, library,
analytics, admin
public/images/ photos (Unsplash licence, see CREDITS.md)
docs/ architecture, design decisions, law sources
LexIntel is an independent decision-support platform, not an official UAE government service, and does not replace licensed legal counsel.