TenderPilot

TenderPilot automates legal operations with AI agents that triage cases, organize evidence, and draft client updates—keeping lawyers in control, not replaced.

Comment

Inspiration

What it does

🧠 TenderPilot — AI Legal Ops Platform

Human-in-the-Loop AI Assistant for Legal Operations

Next.js React TypeScript Supabase Gemini

🚀 Overview

TenderPilot is a next-generation AI platform for legal case management — built using Google’s Agent Development Kit (ADK) principles. It demonstrates advanced multi-agent orchestration, parallel processing, and agent-to-agent (A2A) communication for streamlined legal workflows.

At its core, TenderPilot transforms chaotic client communications into clean, actionable tasks that legal teams can approve and execute with confidence.

⚙️ Key Features

🤖 Multi-Agent Orchestration

  • Evidence Sorter Agent — Extracts and structures billing or medical records.
  • Client Communications Agent — Drafts empathetic and professional client updates.
  • Classifier Agent — Routes new messages to the correct specialists.
  • Orchestrator — Dispatches all AI agents in parallel, tracking every step.

🧍 Human-in-the-Loop Workflow

  • Approval System — Attorneys review AI proposals before execution.
  • Confidence Scoring — Displays the model’s certainty on each task.
  • Execution Preview — See exactly what the AI intends to do before confirming.

📈 Real-Time Monitoring

  • Audit Timeline — Every decision logged for transparency.
  • Telemetry Dashboard — Live metrics for tasks, handoffs, and loop ticks.
  • Task Watcher — Continuous background monitoring for new messages.

🧩 Modern UI / UX

  • Gradient-based design (Blue → Purple)
  • Framer Motion animations
  • Lucide icons + TailwindCSS responsive layout

🧱 ADK Patterns Implemented

Pattern Description How It Works
🔀 Parallelism Multiple agents run concurrently Evidence Sorter + Client Comms execute in parallel
🤝 A2A Communication Agents share knowledge between runs Client Comms imports the latest Evidence Sorter results
🔄 Continuous Loop Background monitoring & auto-classification Watcher pings /api/loop/tick every few seconds

🏗️ System Architecture

┌──────────────┐
│   Inbox      │
└───────┬──────┘
        │
        ▼
┌──────────────┐
│ Classifier   │
└──────┬───────┘
       │
┌──────┴─────────────┐
│  Parallel Agents    │
│  (Evidence & Comms) │
└──────┬──────────────┘
       │
       ▼
┌──────────────┐
│  Approvals   │
└──────┬───────┘
       ▼
┌──────────────┐
│  Execution   │
└──────────────┘

🧩 Tech Stack

Frontend

  • Next.js 16 (App Router)
  • React 19.2
  • TypeScript
  • TailwindCSS v4
  • Framer Motion
  • Lucide React

Backend

  • Next.js API Routes
  • Supabase (PostgreSQL)
  • Google Gemini 2.5 Pro

Infrastructure

  • Vercel (deployment)
  • Supabase Cloud (DB + hosting)

⚡ Getting Started

Prerequisites

Installation

git clone https://github.com/tomiwaaluko/tenderpilot.git
cd tenderpilot
npm install

Environment Setup

Create .env.local:

GEMINI_API_KEY=your_gemini_api_key_here
NEXT_PUBLIC_SUPABASE_URL=your_supabase_url
NEXT_PUBLIC_SUPABASE_ANON_KEY=your_supabase_anon_key
NEXT_PUBLIC_SITE_URL=http://localhost:3000
TP_USE_MOCKS=false

💡 Tip: During testing, set TP_USE_MOCKS=true to use mock Gemini outputs.

Start the Server

npm run dev
# open http://localhost:3000

🧠 Usage Guide

1. Submit a Message → /inbox

Paste a case ID and client message (email/text). TenderPilot classifies and creates structured tasks.

2. Start Watcher → /tasks

Enables the continuous ADK loop. Automatically detects new messages and classifies them.

3. Run Orchestrator

Dispatches agents in parallel (Evidence Sorter + Client Comms). Creates a parallel_dispatch entry in the audit log.

4. Review & Approve → /approvals

Attorneys approve AI-generated proposals. Client Comms uses A2A handoff from Evidence Sorter.

5. Monitor Everything

  • /audit → Action timeline (loop ticks, handoffs, dispatches)
  • /telemetry → Live metrics for ADK activity

🧪 Testing (Quick Validation)

# Create a new message
curl -X POST http://localhost:3000/api/ingest \
  -F "caseId=case-101" \
  -F "text=Client sent new bills from Orlando Health"

# Run orchestrator
curl -X POST http://localhost:3000/api/orchestrator/run

# Simulate watcher loop
curl -X POST http://localhost:3000/api/loop/tick

Check /audit for:

  • parallel_dispatch
  • a2a_handoff
  • loop_tick

📊 Telemetry Metrics

Metric Description
Pending / Proposed / Approved / Executed Task pipeline counts
Parallel Dispatches Number of ADK multi-agent runs
A2A Handoffs Agent-to-agent data transfers
Last Loop Tick Time of last watcher activation

🧾 Project Structure

app/
 ├── api/
 │   ├── agents/
 │   │   ├── client-comms/       # Client communication agent
 │   │   └── evidence-sorter/    # Evidence extraction agent
 │   ├── classify/               # Message classification
 │   ├── ingest/                 # Message ingestion
 │   ├── loop/tick/              # Continuous watcher
 │   └── orchestrator/run/       # Parallel dispatcher
 ├── inbox/                      # Inbox UI
 ├── tasks/                      # Orchestration dashboard
 ├── approvals/                  # Human review UI
 ├── audit/                      # System timeline
 └── telemetry/                  # ADK metrics dashboard

🧰 Troubleshooting

Issue Fix
“Ingest failed” Check Supabase connection + disable RLS
“Gemini API error” Verify GEMINI_API_KEY or use mocks
Tasks missing Ensure classification + orchestrator ran
Watcher inactive Confirm /api/loop/tick returns 200
Telemetry empty Run through one full workflow first

🌐 Deployment (Vercel)

  1. Push to GitHub:
   git push origin main
  1. Import to Vercel
  2. Add .env variables under project settings
  3. Deploy → open https://your-domain.vercel.app

🤝 Contributing

Pull requests are welcome! Follow standard TypeScript + ESLint rules, commit semantically, and document new routes or agents.

🧭 Roadmap

Phase Features Status
1. Core System Classification, Approvals, Telemetry ✅
2. ADK Integration Parallel Agents, A2A, Continuous Loop ✅
3. Future Auth, File Uploads, Analytics, Tests 🚧

🔐 Security

Development:

  • RLS disabled for demo simplicity
  • API keys stored safely in .env.local

Production:

  • Enable RLS
  • Add authentication (NextAuth.js)
  • Enforce HTTPS
  • Rate-limit endpoints

🏆 Credits & Acknowledgments

  • Google Gemini — LLM powering intelligent classification
  • Supabase — Realtime database and backend
  • Vercel — Hosting & deployment
  • Next.js Team — Framework excellence

💡 Built For

Google AI Hackathon 2025 — Showcasing multi-agent ADK architecture in real legal workflows. “Human judgment + AI efficiency = Better justice.”

Last updated: **October 26, 2025**

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for TenderPilot

Built With

  • framer-motion
  • google-agent-development-kit-(adk)
  • google-gemini-2.5-pro-api
  • lucide-react
  • next.js
  • node.js
  • react-19
  • shadcn/ui
  • supabase-(postgresql)
  • tailwind-css
  • typescript
  • vercel
Share this project:

Updates