Past event#San Francisco

Agent Foundry

Overview

Agent Foundry – August 23, 2025

An invite-only, founder-focused sprint where 200 of the brightest developers, researchers, and AI innovators build venture-scale products with AI agents and Model Context Protocol (MCP). Designed for exceptional talent ready to launch the next wave of AI-driven startups.

Selected participants will also have the chance to be considered for Afore Capital’s pre-seed program, investing $500k–$2M+ in exceptional founders.

Theme: Building the Future with AI Agents & MCP
📍 Exact location disclosed upon acceptance
🔗 Learn more: https://www.agentfoundry.org

Tracks

Best use of Composio

Partner track

Best use of Phenoml

Partner track

Best use of Qdrant

Partner track

Projects

Winner

Clara Medical

Project Summary: Medical Bill Audit System "Clara Medical" consists of two main applications designed to help users analyze and negotiate medical bills using AI technology. It makes it easy for patients (and by extension insurance companies) to understand their medical bills and automatically (using an agent) dispute potential discrepancies and mischarges from a crowd-sourced typical cost associated with the procedures done. 🏗️ Architecture Overview Frontend UI (apps/ui/) - Next.js web application Voice Agent Backend (apps/voice-agent/) - Express.js API server 📱 Frontend Application (UI) Technology Stack: Next.js 15 with React 19 TypeScript for type safety Tailwind CSS for styling Radix UI components for accessible UI elements Lucide React for icons Sponsors technologies: OpenAI VAPI Key Features: Medical Bill Upload & Analysis: Users can upload medical bills for AI-powered analysis Bill History Management: Track all analyzed bills with filtering and search capabilities Negotiation Tools: Have an agent write and send emails on your behalf and even call (using VAPI) the medical provider’s billing team. Results Dashboard: View analysis results, cost breakdowns, and savings calculations Modern UI: Clean, professional interface with responsive design Methodology: Has a simple page where a patient can upload a file or dump text of an itemized receipt from a doctors office, this data will be sent to a very simple backend Backend processing utilizes ChatGPT and Phenoml, which can extract the exact line items, finding the related medical codes, matching them to a database with common procedure costs (within high and low ranges, think google flights where it shows "this flight is lower than usual" on a simple color coded progress bar), insurance utilization (how much does my insurance provide for this code, and how much did the office charge, this would be overlaid with the procedure costs progress bar). this data will be returned to the frontend in a table for analysis. If relevant, certain line items may be flagged — is the line item overcharged and should be negotiated. We will also flag if a code was incorrectly assigned to a particular line item, and what that ideal code should be. GitHub: https://github.com/rak3rman/audit

RCR3 members

Winner

Harmonia

What it does Harmonia is the first MCP-native AI Music Agent — a one-stop companion for music creation, personalized listening, and agentic tasks. Users can voice-trigger song ideas, remix tracks via provider-agnostic tools, generate covers with BytePlus, enjoy a TikTok-style feed personalized by Qdrant, and get notes/emails or concert reminders through Composio integrations. How we built it We combined Morph LLM for reasoning/planning, Qdrant for personalization memory, Composio for Gmail/Calendar, and BytePlus for image generation. MCP glues these tools together into a flexible agentic workflow. Challenges Integrating multiple sponsor APIs, aligning MCP schemas, handling Gmail OAuth, and stitching a smooth demo under time pressure. Accomplishments Delivered an end-to-end agent demo (create → listen → chat → tasks). First hackathon version with real tool orchestration. What we learned How to design MCP schemas, chain tools into workflows, and balance UX with technical limits. What’s next Expand creator monetization, refine personalization loops, and integrate more MCP tools for a truly AI-native music ecosystem.

RY2 members

Winner

MediMentor

Inspiration "Health is wealth" — a phrase we’ve all heard, yet the true weight of it only struck us when we lost loved ones during the pandemic and to other health conditions. In those difficult moments, the biggest frustration wasn’t just the loss — it was not knowing why. Not understanding the cause, the condition, or what could have been done differently. This pain, this confusion, and this helplessness became personal to us. And from that place, the idea for MEDIMENTOR was born. What it does We built an AI-powered engine designed to bridge the gap between complex medical information and everyday understanding. MEDIMENTOR takes ambiguous, technical lab results and transforms them into clear, easy-to-understand explanations. Use Case 1: Translating Lab Results into Plain Language Medical lab reports are often filled with complex, technical terms that can be difficult to interpret. MEDIMENTOR takes those ambiguous lab results and explains them in simple, easy-to-understand language — just like a friend would. We ensure that users not only know what their results mean but also understand how it might affect their health. Use Case 2: Recommending the Right Healthcare Provider Understanding your results is only the first step. MEDIMENTOR goes a step further by identifying the type of specialist or provider a user may need to consult based on their lab results. For example, if your HbA1c level is high, MEDIMENTOR will explain what that means and recommend consulting an endocrinologist, allowing you to directly schedule an appointment from the platform. This makes the journey from diagnosis to treatment faster, clearer, and more connected. How we built it The website frontend was made using HTML, CSS, Bootstrap and JS. The website backend was made using Java and Springboot. Gihub repo can be found on https://github.com/saloni0104/MediMentor We also used FHIR API from PhenoML mainly Design was done using Canva Overall, it was a great learning experience trying our hands on several new software and building the site from scratch during the weekend. Challenges we ran into Coming up with the idea was not the difficult part but navigating the execution was. Accomplishments that we're proud of We were able to successfully leverage PhenoML technologies and integrate them in our project. We are proud to have worked on a problem that impacts millions of lives in the field of healthcare everyday. Through this project, we stepped into the shoes of patients and understood the niche pain points that they have in understanding basic medical needs like understanding lab reports in common terminology and identifying the correct provider based on symptoms and emotions. What we learned We learnt to use PhenoML product and different APIs that PhenoML offers. We learnt to use Springboot and integration with PhenoML for our usecas What's next for MediMentor Improve Appointment discovery. Handwritten prescriptions converted into Printed readable format. Knowing the correct prescriptions for general and branded medicines.

PRS3 members

Winner

VibeDebugger

VibeDebugger 🚀 A powerful AI-powered codebase exploration and debugging tool that helps developers understand, analyze, and interact with their repositories through semantic search and intelligent chat assistance. ✨ Features 🔍 Semantic Code Search Vector-based similarity search using OpenAI embeddings Find code by meaning, not just keywords Precise file location and line number results Support for 30+ programming languages 🤖 AI-Powered Chat Assistant Context-aware conversations about your codebase Code explanation and debugging suggestions Architecture insights and refactoring recommendations Integrated with repository knowledge 🎤 Voice Assistant Integration VAPI-powered voice interactions Hands-free codebase exploration Voice-activated feature introductions Real-time conversation capabilities 📊 Repository Analytics Comprehensive repository analytics showing commit history, file changes, and contributor data visualization. The dashboard displays detailed insights into code activity patterns, contribution metrics, and repository evolution over time. AI-generated code detection analysis Commit timeline visualization with contributor insights Repository health and complexity metrics Interactive charts and statistics 🔄 Smart Repository Management GitHub OAuth integration Automated code indexing with progress tracking Support for both public and private repositories Optimized handling of large repositories (500+ files) 🎨 Modern UI/UX Clean, responsive dashboard interface Syntax highlighting for code blocks Clickable file paths with GitHub integration Real-time status updates and notifications 🛠️ Tech Stack Frontend Next.js 15.5 - React framework with App Router TypeScript - Type-safe development Tailwind CSS 4.0 - Utility-first styling Lucide React - Beautiful icons React Markdown - Rich text rendering Backend NextAuth.js 5.0 - Authentication with GitHub OAuth Prisma - Database ORM Neon DB - Serverless PostgreSQL database platform PostgreSQL - Primary database (via Neon DB) Qdrant - Vector database for semantic search AI & ML OpenAI GPT-4 - Chat completions and analysis OpenAI text-embedding-ada-002 - Text embeddings VAPI - Voice assistant integration Custom AI Detection - Code analysis algorithms Infrastructure Docker - Containerized services GitHub API - Repository data fetching Octokit - GitHub API client 🚀 Quick Start Prerequisites Node.js 18+ and npm Neon DB account (serverless PostgreSQL) Docker and Docker Compose (for Qdrant) GitHub OAuth App credentials OpenAI API key VAPI account (optional, for voice features) Installation Clone the repository git clone https://github.com/your-username/vibe-debugger.git cd vibe-debugger Install dependencies npm install Set up environment variables Create a .env.local file: # NextAuth.js NEXTAUTH_SECRET=your-secret-key NEXTAUTH_URL=http://localhost:3000 # GitHub OAuth GITHUB_CLIENT_ID=your-github-client-id GITHUB_CLIENT_SECRET=your-github-client-secret # Database (Neon DB Connection String) DATABASE_URL=postgresql://username:password@ep-example-123.us-east-1.aws.neon.tech/vibe_debugger?sslmode=require # OpenAI OPENAI_API_KEY=sk-your-openai-api-key # Qdrant Vector Database QDRANT_URL=http://localhost:6333 # VAPI Voice Assistant (optional) NEXT_PUBLIC_VAPI_PUBLIC_KEY=your-vapi-public-key 4. **Start Qdrant with Docker** ![Qdrant Running on Docker](imgs/qdrant_running_on_docker.jpg) *Qdrant vector database running successfully on Docker, showing the container status, port mapping, and storage configuration. This setup provides the vector search capabilities essential for semantic code search.* ```bash docker run -p 6333:6333 -v $(pwd)/qdrant_storage:/qdrant/storage qdrant/qdrant Set up the database # Deploy migrations to Neon DB npx prisma migrate deploy npx prisma generate Start the development server npm run dev Open your browser Navigate to http://localhost:3000 📖 Usage Guide Getting Started Sign in with GitHub - Authenticate using your GitHub account Add a Repository - Click "Add Repository" and enter a GitHub URL Index the Repository - Wait for the AI to process and index your code Start Exploring - Use chat or search to interact with your codebase Key Workflows 🔍 Semantic Search "Where is the user authentication logic?" "Find functions that handle file uploads" "Show me error handling patterns" 💬 AI Chat Examples "Explain how this authentication system works" "What are the main security vulnerabilities?" "How can I optimize the database queries?" "Suggest refactoring opportunities" 🎤 Voice Commands "Tell me about this repository" "What are the main features?" "How do I get started with development?" 🏗️ Architecture System Overview High-level system architecture showing the interaction between frontend components, backend services, and external APIs. The diagram illustrates how VibeDebugger integrates Next.js, PostgreSQL, Qdrant vector database, and various AI services. ┌─────────────────┐ ┌─────────────────┐ ┌─────────────────┐ │ Next.js App │◄──►│ PostgreSQL │ │ Qdrant │ │ (Frontend) │ │ (Metadata) │ │ (Embeddings) │ └─────────────────┘ └─────────────────┘ └─────────────────┘ │ │ │ ▼ ▼ ▼ ┌─────────────────┐ ┌─────────────────┐ ┌─────────────────┐ │ GitHub API │ │ OpenAI API │ │ VAPI Service │ │ (Repository) │ │ (Embeddings/Chat)│ │ (Voice) │ └─────────────────┘ └─────────────────┘ └─────────────────┘ Data Flow Detailed sequence diagram showing the complete workflow from repository indexing to AI-powered code search and chat interactions. This diagram demonstrates how user queries are processed through vector embeddings and semantic search to provide intelligent responses. Repository Ingestion: GitHub API → Code Parsing → Chunking Vector Generation: Code Chunks → OpenAI Embeddings → Qdrant Storage Semantic Search: Query → Embedding → Vector Search → Results AI Chat: Context + Query → GPT-4 → Response + Code References Key Components /src/lib/code-indexer-optimized.ts - Smart repository indexing /src/lib/ai-detector.ts - AI-generated code analysis /src/components/chat/ - Chat interface and message handling /src/components/voice-assistant/ - Voice interaction components /src/app/api/ - Backend API routes 🎯 AI Detection Features VibeDebugger includes advanced AI detection capabilities: Detection Methods Perplexity Analysis - Statistical language patterns Syntax Perfection Scoring - Code consistency metrics Comment Quality Assessment - Documentation patterns Structural Analysis - Code organization patterns Pattern Recognition - Common AI-generated signatures Accuracy Disclaimer Results are based on 2024 AI detection research and are probabilistic, not definitive. Use as guidance rather than absolute determination. 🔧 Configuration Environment Variables Variable Description Required Default NEXTAUTH_SECRET NextAuth.js encryption key ✅ - NEXTAUTH_URL Application base URL ✅ - GITHUB_CLIENT_ID GitHub OAuth app ID ✅ - GITHUB_CLIENT_SECRET GitHub OAuth app secret ✅ - DATABASE_URL Neon DB PostgreSQL connection string ✅ - OPENAI_API_KEY OpenAI API key ✅ - QDRANT_URL Qdrant vector database URL ✅ http://localhost:6333 NEXT_PUBLIC_VAPI_PUBLIC_KEY VAPI voice service key ❌ - GitHub OAuth Setup Go to GitHub → Settings → Developer settings → OAuth Apps Create a new OAuth App with: Homepage URL: http://localhost:3000 Authorization callback URL: http://localhost:3000/api/auth/callback/github Copy the Client ID and Client Secret to your .env.local OpenAI API Setup Visit OpenAI Platform Create an API key with access to: text-embedding-ada-002 model gpt-4 model (or gpt-3.5-turbo as fallback) Add the API key to your environment variables Neon DB Setup Neon DB dashboard showing database statistics, contributor data analytics, and performance metrics. The dashboard provides real-time insights into database usage, query performance, and contributor activity patterns. Create a Neon DB account Create a new project and database Copy the connection string from your Neon dashboard The connection string format: postgresql://username:password@ep-xxx-xxx.region.aws.neon.tech/database_name?sslmode=require Add the connection string to your .env.local as DATABASE_URL Neon DB Benefits: Serverless PostgreSQL with automatic scaling Built-in connection pooling Branch-based development workflows No database maintenance required 📊 Performance & Limits Repository Size Handling Small repos (< 100 files): Full indexing Medium repos (100-800 files): Smart filtering Large repos (> 800 files): Priority-based selection (500 files max) Rate Limiting GitHub API: 5,000 requests/hour (authenticated) OpenAI API: 3 requests/second for embeddings Batch Processing: 3 files per batch with 1s delays Storage Requirements Neon DB: ~1KB per repository metadata (serverless PostgreSQL) Qdrant: ~6KB per code chunk (1536-dimension vectors) Estimated: 50MB vector storage per 10,000 lines of code 🚀 Deployment Production Environment Database Setup bash # Deploy migrations to production Neon DB npx prisma migrate deploy Production Neon DB Configuration: Use production branch in Neon console Enable connection pooling for better performance Set up database branching for staging environments Configure automatic backups and point-in-time recovery Qdrant Deployment # Using Docker in production docker run -d --name qdrant-prod \ -p 6333:6333 \ -v /path/to/storage:/qdrant/storage \ qdrant/qdrant Build and Start npm run build npm start Docker Deployment # Dockerfile example FROM node:18-alpine WORKDIR /app COPY package*.json ./ RUN npm ci --only=production COPY . . RUN npm run build EXPOSE 3000 CMD ["npm", "start"] 🧪 Testing Run Tests # Run all tests npm test # Run with coverage npm run test:coverage # Run specific test file npm test -- --testPathPattern=ai-detector Manual Testing Checklist [ ] GitHub OAuth authentication [ ] Repository addition and indexing [ ] Semantic search functionality [ ] AI chat responses [ ] Voice assistant (if enabled) [ ] Repository deletion [ ] Error handling and recovery 🤝 Contributing We welcome contributions! Please see our Contributing Guide for details. Development Workflow Fork the repository Create a feature branch (git checkout -b feature/amazing-feature) Make your changes Run tests (npm test) Commit your changes (git commit -m 'Add amazing feature') Push to the branch (git push origin feature/amazing-feature) Open a Pull Request Code Style TypeScript: Strict type checking enabled ESLint: Extended Next.js configuration Prettier: Automated code formatting Conventional Commits: Standardized commit messages 📝 API Reference Repository Management POST /api/repositories - Add new repository GET /api/repositories - List user repositories POST /api/repositories/[id]/index - Start indexing POST /api/repositories/[id]/cancel-index - Cancel indexing DELETE /api/repositories/[id]/delete - Delete repository Search & Chat POST /api/search - Semantic code search POST /api/chat - AI chat completions GET /api/repositories/[id]/commits - Repository commits 🐛 Troubleshooting Common Issues 1. "Qdrant connection failed" # Check if Qdrant is running docker ps | grep qdrant # Start Qdrant if not running docker run -p 6333:6333 qdrant/qdrant 2. "OpenAI API rate limit exceeded" Check your OpenAI account usage and billing Reduce batch sizes in indexing configuration Implement request queuing for large repositories 3. "GitHub OAuth callback error" Verify callback URL matches OAuth app configuration Check NEXTAUTH_URL environment variable Ensure GitHub OAuth app is active 4. "Database migration failed" # For development (local database reset) npx prisma migrate reset npx prisma migrate dev # For production (Neon DB) npx prisma migrate deploy 5. "Neon DB connection issues" Verify your connection string includes ?sslmode=require Check if your Neon project is active (not suspended) Ensure your IP is whitelisted if using IP restrictions Try connecting directly from Neon console to test connectivity Debug Mode Enable detailed logging: NODE_ENV=development DEBUG=true 📄 License This project is licensed under the MIT License - see the LICENSE file for details. 🙏 Acknowledgments OpenAI for GPT-4 and embedding models Qdrant for vector database technology GitHub for repository API and OAuth VAPI for voice assistant capabilities Vercel for Next.js framework and hosting platform 📧 Support Documentation: GitHub Wiki Issues: GitHub Issues Discussions: GitHub Discussions Email: support@vibedebuger.com Built with ❤️ by developers, for developers Making code exploration intelligent and intuitive

XA2 members

Winner

Medivice

Inspiration The medical industry still relies heavily on legacy systems that often slow down processes for both patients and doctors. We saw an opportunity to modernize patient intake by reducing repetitive tasks and freeing up doctors’ time. Our idea began with figuring out how we could save time for doctors. We realized one of the ways we can do this is by automating the collection of preliminary patient information such as personal details, symptoms, and the reason for their visit. What it does Medivice allows patients to call an AI assistant that will gather important patient intake information and record the data. Then the application provides a dashboard for doctors to utilize and view updates from patients about their most urgent issues, symptoms, and conditions, and how they could be addressed. How we built it We used OpenAI, Vapi, Phenoml, Deepgram, and FastAPI for the backend, and NextJS for the frontend patient information dashboard. Challenges we ran into Integrating with Vapi and other APIs Debugging null errors and handling edge cases with AI Balancing speed of development with system reliability Accomplishments that we're proud of Building a working prototype in just a few hours Creating a functional pipeline from patient call → AI intake → doctor dashboard Demonstrating how AI can meaningfully save doctors time What we learned AI has enormous potential in healthcare and is very cool, but humans are still necessary. There's a lot that AI can do and it's pretty cool. Building Medivice made us realize that the future lies in continued collaboration between AI and humans. What's next for Medivice Enhancing patient intake with more structured and customizable forms Expanding voice-based intake for smoother patient experiences Adding data export options for integration with other EHR systems

EJ2 members

Winner

Quorix

Checkout the ReadMe for technical details Problem: Leaders report >10% denial rates are common; missing/inaccurate data and prior auth drive preventable denials. Claim denial rates vary 1%–54% by insurer. Doctors DON’T like spending their time optimizing claims for payouts. Existing Solutions: Scribes document appointments, can’t create codes and claims. EHR auto-coding is generic and not specific enough for many insurance providers RCM platforms catch basics, yet initial denials still hover ~11–12% — payer-specific nuances slip through. Quorix An application that extracts ICD‑10 from notes, asks targeted clarifiers (laterality, specificity, sequencing), and validates against payer‑specific rules (Anthem, UnitedHealthcare, etc.) before submission — turning ambiguous documentation into payer‑aligned, first‑pass clean claims. Market Potential: RCM market hundreds of billions and growing at double‑digit CAGR — budget exists for denial prevention. Medical coding market >$14B with ~10% CAGR — sustained demand for accurate ICD capture. AI in RCM projected multi‑x growth (20B→180B decade) — strong tailwinds for real‑time coding intelligence. Future Vision Scale payer logic nationwide; close the 11–19% denial gap with real‑time, payer‑specific clarification. Reduce preventable denials, slash appeals, return time to patients. Migrate to Chrome Extension for EHR compatibility

SS2 members

See all projects