When people in Korea face legal issues, they have three bad options: expensive lawyers ($75+ per session), unreliable internet searches, or AI chatbots that hallucinate laws that don't exist.
I built Lawmadi OS to fix this — an AI legal operating system with 60 domain-specialized agents that verify every answer against live government databases.
Live: lawmadi.com
The Problem with Legal AI
Ask ChatGPT about Korean labor law, and it will confidently cite "Article 27 of the Labor Standards Act" — except that article might not say what it claims, or might not exist at all. In the legal domain, hallucination isn't just annoying — it's dangerous. People make life-changing decisions based on legal information.
How Lawmadi OS Works
3-Layer NLU Routing
Instead of sending every query through an expensive LLM classification step, we use cascading routing:
This gives us low latency for most queries, high accuracy (264/264 test cases passing), and cost efficiency.
60 Specialized Agents
Each of the 60 agents specializes in a specific area of Korean law:
- L09 담우 — Labor Law (unfair dismissal, unpaid wages)
- L08 온유 — Lease/Rent Law (전세 deposits, tenant rights)
- L03 담슬 — Divorce & Family Law
- L10 결휘 — Traffic Accidents
- L01 휘율 — Criminal Law
- And 55 more covering tax, IP, immigration, inheritance, medical, military, environment, data privacy, startups, etc.
Why 60 instead of 1 generalist? Specialization matters. Each agent has domain-tuned prompts, knowledge of relevant statutes, and optimized response patterns. It's like having a law firm with 60 specialists instead of one generalist.
4-Stage Verification Pipeline
This is the core architecture:
Stage 4 is what makes Lawmadi OS different. After Gemini generates a response, we:
- Extract all statute citations from the response
- Query Korea's official legislative database (법제처, law.go.kr) via DRF API
- Verify — Does the law exist? Does the article number exist? Is the content accurate?
- Score — Generate a 0-100 verification score
- Decide — If score is below threshold, reject the entire response
We cross-reference against 10 government data sources:
- Statutes (법령)
- Enforcement Decrees (시행령)
- Enforcement Rules (시행규칙)
- Court Precedents (판례)
- Administrative Rules (행정규칙)
- And 5 more
Fail-Closed Design
If the verification API is down, the system doesn't fall back to unverified responses. Instead:
- Circuit breaker trips after consecutive failures
- System enters fail-closed mode
- All responses are held until verification is available
- We'd rather give no answer than an unverified one
The 5-Stage Empathy Framework
Legal issues are stressful. Every response follows this structure:
- Emotional acknowledgment — "This situation must be frustrating..."
- Situation diagnosis — Clear analysis of the legal issue
- Action roadmap — Specific steps with deadlines
- Safety net — Legal aid resources, hotlines, government services
- Supportive closing — Encouragement and next steps
Results After 1 Week
| Metric | Value |
|---|---|
| Unique Visitors | 114 |
| Queries Processed | 481 |
| Success Rate | 99.6% |
| Avg Verification Score | 84.7/100 |
| Korean Citation Accuracy | 82.5% |
| English Citation Accuracy | 25.6% (improving) |
| Tests Passing | 282/282 |
| Avg Response Time | ~40s |
Most popular domains: Labor law (90 queries), Housing/Lease (83), Divorce (50), Traffic accidents (48)
Tech Stack
| Component | Technology |
|---|---|
| Backend | FastAPI 0.128.0 + Python 3.10+ |
| LLM | Google Gemini 2.5 Flash |
| RAG | Vertex AI Search (14,601 docs) |
| Verification | 법제처 DRF API (10 SSOT sources) |
| Database | Cloud SQL PostgreSQL 17 |
| Hosting | GCP Cloud Run + Firebase |
| Billing | Paddle (credit-based) |
| CI/CD | GitHub Actions (5-stage pipeline) |
| Auth | JWT RBAC + Email OTP |
| Anti-abuse | IP + Canvas Fingerprint + Device Token |
Pricing
- Free: 2 queries/day (no account needed)
- Starter: 20 queries — .50
- Standard: 100 queries — .99
- Pro: 300 queries — .99
Credit-based, no subscription. Powered by Paddle.
Challenges & Next Steps
- Latency — ~40s avg is too slow. Gemini generation (~30s) is the bottleneck. Exploring parallel RAG + prefetch.
- English citations — 25.6% accuracy vs 82.5% Korean. Standardized English translations of Korean law names are inconsistent.
- Scale — 60 system prompts to maintain. Considering automated prompt generation.
Try It
- Korean: lawmadi.com
- English: lawmadi.com/en
- 60 Leaders: lawmadi-db.web.app/leaders
I'd love to hear your thoughts, especially on:
- Multi-agent specialization vs. single generalist approaches
- Fail-closed verification in AI systems
- Ideas for reducing response latency
Built by Jainam Choe — choepeter@outlook.kr
