Hiromi

💡 Inspiration

The U.S. criminal justice system processes millions of cases annually, yet many convicted individuals lack access to a thorough legal analysis that could reveal grounds for appeal. Wrongful convictions and procedural errors often go unnoticed due to:

Long waiting time because of manual review of witness testimony, physical evidence, and documentary evidence
Limited public defender resources and overwhelming caseloads
Bias and inconsistencies in original trials that may be overlooked

🔍 What It Does

Hiromi is an intelligent multi-agent system that performs comprehensive criminal appeal analysis by:

🤖 Multi-Agent Evidence Analysis

Parallel Evidence Processing: Three specialized AI agents analyze different evidence types simultaneously:
- Witness Analysis Agent: Examines eyewitness testimony for bias, inconsistencies, visibility issues, timeline discrepancies, and stress factors
- Physical Evidence Agent: Reviews forensic evidence, chain of custody, and physical proof
- Documentary Evidence Agent: Analyzes court documents, records, and written evidence
Executor Agent Debate System: Three executive agents debate the findings:
- Each executor specializes in their evidence domain
- Agents must reach consensus (unanimous 3/3 or majority 2/3)
- Built-in retry mechanism ensures thorough deliberation
- Offers a preliminary recommendation: INNOCENT (mistrial advised) or NOT INNOCENT, pending human validation.
Automated Legal Memo Generation:
- Generates comprehensive PDF memos with verdict justification
- Cites specific evidence from all three analysis streams
- Uploads memos to Google Cloud Storage for attorney access

📧 Gmail Integration

Automated email notifications to attorneys
PDF attachment support for authorization forms
OAuth2 secure authentication

🔄 Orchestration & Scalability

Google Agent Development Kit (ADK) powers the multi-agent architecture
Parallel processing for faster case analysis
Sequential debate coordination with loop retry mechanisms
Session management for a consistent state across agent interactions

🏗️ How We Built It

AI Framework: Google ADK orchestrates our 7-agent pipeline powered by Gemini 2.0 Flash. LiteLLM handles model routing while A2A SDK enables agent communication.

Backend: FastAPI REST API with PostgreSQL on Google Cloud SQL, using SQLAlchemy ORM and Cloud SQL Python Connector for secure connections.

Cloud Services: Google Cloud Storage for PDF memos, Gmail API for email automation.

Pipeline Architecture:

Evidence Collection → Parallel Analysis (3 agents) → Bridge Agent 
→ Executor Debate (3 agents) → Consensus Checker → Final Decision 
→ Memo Generation & Storage

🚧 Challenges We Ran Into

Simulating Lawmatic's dashboard using GitHub.
Implementing the multi-agent system into existing software like GitHub Project (Lawmatic Simulation).
Configuring Google Cloud services into a single ecosystem.

🏆 Accomplishments That We're Proud Of

✅ Successfully implemented a multi-agent pipeline with parallel processing and debate coordination

✅ Built a production-ready REST API with PostgreSQL database and Google Cloud integration

✅ Achieved automated legal analysis that would typically require months of attorney time

✅ Integrated email automation for seamless attorney communication

✅ Designed bias detection algorithms for witness testimony analysis (gender, racial, situational bias)

✅ Built secure cloud infrastructure with proper authentication and data encryption

What we learned

Learned how to create a multi-agent pipeline utilizing Google's Adk
How to utilize Google Cloud service within our multi-agent ecosystem
Converting our multi-agent into existing software(Github Project)

What's next for Hiromi

Integration with Real Case Databases
Expand beyond English-language transcripts and evidence
Collaborate with public defender organizations and legal aid nonprofits to test Hiromi on anonymized case datasets