Award-winning ML project
SimplyLaw
SimplyLaw tackles the real problem facing law firms like Morgan & Morgan: lawyers drowning in messy data - overflowing inboxes, incomplete client texts, and transcripts full of 'uhh.' This award-winning system uses four collaborative AI agents that don't just make API calls - they genuinely work together, sharing context and building on each other's work. **The Four Agents:** • **AI Orchestrator** - Routes requests using keyword scoring and triggers Collaborative Mode when needed • **Doc Agent** - Processes ANY file type (even blurry faxes and password-protected PDFs), extracts key legal info, sticks to facts and logic • **Sherlock Agent** - Finds patterns, builds case timelines, evaluates settlements, brings creative strategic thinking • **Client Comms Agent** - Drafts legally-safe messages in plain English **What Makes It Unique:** True agent collaboration where agents debate and reach consensus (max 10 iterations) before delivering responses. The system handles real-world chaos: blurry scanned faxes, poor-quality audio, handwritten notes - with robust error handling and multiple processing fallbacks. **Production Features:** Complete FastAPI server with 8 REST endpoints, Twilio integration for client updates, Resend email dispatch from custom domain, **RAG** for document search and strategy generation, deployed on AWS infrastructure with Docker.
Problem
Legal teams lose time and context when case information arrives as inconsistent PDFs, scans, audio, faxes, and client messages.
My role
Developed the multi-agent workflow and AI Orchestrator with intent classification, keyword scoring, and agent routing, then deployed the backend with AWS and Docker.
Approach
Coordinate four specialized Google ADK agents for intake, document analysis, case reasoning, and legally careful client communication behind a FastAPI service.
Outcome
Won 1st Place in the Morgan & Morgan challenge at KnightHacks VIII with a working legal-automation system.
Implementation
- Built Orchestrator, Document, Sherlock, and Client Communications agent roles.
- Combined OCR, document parsing, RAG, REST endpoints, and collaborative agent turns.
- Packaged and deployed the backend with Docker on AWS.
Evaluation
- Exercised routing across communication, document-processing, and strategy requests.
- Tested extraction fallbacks against heterogeneous file inputs.
- Limited collaborative debates to a deterministic maximum of ten iterations.
Results
- Delivered a functioning multi-agent legal workflow spanning intake, analysis, and communication.
- Awarded 1st Place in the Morgan & Morgan challenge at KnightHacks VIII.
Constraints
- Inputs included low-quality scans, audio, handwriting, and protected PDFs.
- Agent outputs needed traceable routing and conservative client-facing language.
- The service needed fallbacks when an extraction path failed.
Tradeoffs
- Specialized agents improve separation of concerns but increase orchestration complexity.
- Generated analysis supports legal teams and requires professional review.
Next improvements
- Add stronger citation provenance and case-level evaluation sets.
- Expand permission controls and audit trails for production legal environments.