The problem LaborLex-AI solves

Democratizing Legal Protection with Pay-Per-Prompt Access

• Millions of Indonesian job seekers sign "toxic" employment contracts because professional legal review is prohibitively expensive, and existing AI legal tools often require expensive monthly subscriptions that don't make sense for a one-time user.

• LaborLex AI bridges this gap as a Micro-SaaS Legal Assistant specifically tuned for Indonesian Labor Laws (Omnibus Law).

Pay-Per-Prompt Accessibility: Instead of forcing users into expensive subscriptions, we implement a Pay-Per-Prompt model. Users pay a micro-fee (via Visa/Digital Payment) only when they need a contract reviewed. This is powered by an x402 (HTTP 402 Payment Required) architecture, making high-quality legal analysis accessible to fresh graduates and blue-collar workers.

RAG-Powered Accuracy: The system uses Retrieval Augmented Generation to cross-reference uploaded contracts against the latest UU Cipta Kerja and PP Jaminan Sosial, detecting predatory clauses like unlawful penalty fees or hidden "internship" traps.

Detecting "Toxic" Clauses: It instantly identifies red flags, such as "salary below minimum wage," "illegal holding of diplomas (ijazah)," or "discriminatory termination clauses," explaining why they are illegal in simple language.

Cost-Effective Architecture: By optimizing the token usage and RAG retrieval, we provide "Tier-1 Law Firm" insights at the price of a cup of coffee.

Empowerment: It transforms a daunting legal document into actionable insights, giving workers the confidence to negotiate or decline exploitative offers.

Challenges we ran into

1. The "Dirty Data" Nightmare: Scanned PDFs vs. RAG One of the biggest hurdles wasn't the code, but the data itself. The Indonesian Omnibus Law documents available weren't clean digital text files; they were scanned images of printed papers (some seemingly taken with a mobile phone camera) converted into PDF.

• The Hurdle: Standard PDF loaders couldn't extract text for the RAG system, making chunking impossible.

• The Struggle: I spent significant time trying to clean this unstructured data so the AI could actually "read" the laws instead of seeing blank pages.

2. Prompt Engineering & System Behavior Getting the AI to act as a strict legal assistant was difficult. Early on, the model would hallucinate or be too conversational. Crafting the perfect System Prompt to ensure the AI strictly adheres to the provided legal context without making up facts required countless iterations of trial and error.

3. Zero-Budget Development: From Local PC to Alibaba Cloud As a developer with limited resources, I had to be strategic about costs:

• Local Beginnings: During development, I relied entirely on LM Studio running on my local PC to save money. I utilized text-embedding-qwen3-embedding-8b for vectorization and openai/gpt-oss-20b for inference. This allowed me to iterate fast without burning cash.

• The Switch: Now, for better performance, I've migrated to Alibaba Cloud's Free Tier, utilizing the qwen3-max-2026-01-23 model (which offers a generous 1M token grant). Managing this transition from a local environment to a cloud provider while keeping the system stable was a steep learning curve.

4. The Two-Week Sprint: Mastering x402 Perhaps the most personal challenge was the timeline. I was only introduced to the x402 (Payment) protocol less than two weeks ago during the Base Workshop in Jogja.

• We had to learn a completely new monetization standard and deliver a working mini-app in a very short time.

• While this app might look simple to others, for us, shipping a functional RAG system integrated with x402 payment logic in such a tight window is a massive personal