Building an AI Contract Review System with Next.js and LLMs

Contracts are full of legal jargon, hidden risks, and ambiguous clauses that often require manual review. During the development of ContractIQ, I wanted to explore how modern AI models could assist professionals in reviewing contracts faster while maintaining human oversight.

The result was an AI-powered contract review platform capable of extracting clauses, identifying risks, generating recommendations, and producing professional review reports.

The Problem

Traditional contract review is:

• Time-consuming
• Prone to human oversight
• Difficult to scale across large volumes of documents

Many organizations spend hours reviewing agreements to identify:

• Risky clauses
• Missing obligations
• Ambiguous language
• Compliance concerns

This made contract analysis a perfect use case for AI.

System Architecture

The platform is built using:

• Next.js
• TypeScript
• PostgreSQL
• Prisma
• Trigger.dev
• GPT-OSS-120B
• Gemini 2.5 Flash

The workflow follows a simple pipeline:

1. Upload contract PDF
2. Extract text and clauses
3. Analyze each clause with AI
4. Generate risk assessments
5. Produce recommendations
6. Create a professional review report

Why Multiple AI Models?

Instead of relying on a single model, I experimented with multiple LLMs.

GPT-OSS-120B

Used for:

• Deep clause analysis
• Risk detection
• Legal reasoning
• Recommendation generation

Gemini 2.5 Flash

Used for:

• Fast summaries
• Classification tasks
• Workflow acceleration

This hybrid approach improved both speed and output quality.

Human-in-the-Loop Review

One challenge with AI systems is trust.

Even the best models can hallucinate or misinterpret legal language.

To address this, I implemented a human-in-the-loop workflow where reviewers can:

• Approve AI suggestions
• Reject inaccurate assessments
• Add manual annotations
• Finalize reports before delivery

This keeps humans in control while benefiting from AI automation.

Real-Time Processing

Contract analysis can take time, especially for large documents.

To improve user experience, I implemented:

• Real-time summary streaming
• Background job processing
• Durable workflows with Trigger.dev
• Automated email delivery

Users receive updates without waiting for long-running tasks to complete.

Key Lessons Learned

Building ContractIQ taught me several valuable engineering lessons:

• AI products require strong workflow orchestration.
• Human oversight is critical for high-stakes applications.
• Prompt engineering alone is not enough; system design matters.
• Streaming responses significantly improve perceived performance.
• Reliability becomes more important as AI complexity increases.

Future Improvements

Some features I plan to explore:

• Clause similarity search using vector databases
• RAG-based legal knowledge retrieval
• Multi-document comparison
• Team collaboration workflows
• Compliance scoring dashboards

Conclusion

Building ContractIQ was an excellent opportunity to combine full-stack development, workflow orchestration, and large language models into a practical real-world application.

The project reinforced an important lesson:

As AI systems become more capable, designing reliable workflows and maintaining human oversight will remain essential for building trustworthy software.