Jay Ozer

PoppyNote was born from a problem hitting close to home: my wife Andrea is a pediatric dentist, and she'd routinely stay late at the office finishing up her clinical notes after a long day with patients. It was the administrative task she complained about most. I searched for tools we could buy, but the cost didn't justify the value—plus, everything on the market was built for all health professionals, not specifically for dentistry. I wanted something dental-specific that could be fine-tuned with custom instructions for pediatric workflows. So I built PoppyNote with HIPAA-ready AWS infrastructure (S3, Lambda, CloudWatch with encryption at rest). The system records patient visits and automatically converts them into structured SOAP notes. But I didn't stop there—I built a full integration with Curve, the dental management software Andrea uses. PoppyNote pulls today's patient list via Curve's API, creates SOAP notes from the visit recordings, and automatically attaches them back to the notes section in Curve. Now Andrea doesn't have to stay late writing notes, and the documentation captures everything mentioned during the visit without her having to remember details hours later. The real game-changer is the chat function: using Pinecone Assistant with a RAG approach, clinicians can chat with prior SOAP notes before a visit—either getting a summary or asking specific questions about a patient's history. We're now expanding beyond our practice to serve all dental professionals, starting with orthodontists, and I'm actively working on integrations for other dental management software platforms.

I built this because why not. Everything was MCP this, MCP that—so I thought, what if I had my own MCP server that could do tool calls to answer questions about me? MCP jozer is exactly that: a Model Context Protocol server that provides comprehensive personal information through 14 standardized tools accessible to any MCP-compatible client. Built with FastMCP, the server exposes tools for biography, professional experience, skills, education, projects, contact information, and CV metadata. This implementation demonstrates how to build personal API endpoints following the emerging MCP standard for AI agent integration. The server can be deployed to FastMCP Cloud or self-hosted, enabling AI assistants like Claude Desktop to access structured personal data through a unified protocol. It's open source, took me under an hour to build (as documented in my Medium post), and honestly, it's just fun to have your own AI-accessible API for yourself.

IntentIQ was born from an operational crisis when interest rate drops triggered a massive surge in order volume, overwhelming the operations team with 14,000+ emails daily. The system implements a sophisticated two-stage AI pipeline: first, a triage model classifies emails as actionable or unactionable (filtering spam and FYI messages), then actionable emails flow through a category model that routes them into 7 distinct workflow categories including title & curative, claims, loan amount updates, and scheduling. Built with a vendored email parser that extracts the latest message from email threads plus attachments, IntentIQ provides both SDK and CLI interfaces for flexible integration. The platform includes automated acknowledgement responses for specific email types, which spawned the creation of the Agentbuilder Outlook MCP server for seamless Outlook integration.

I'm always looking for ways to make life easier at my wife's pediatric dentistry. We've built a bunch of cool tools—an automated patient intake form, the Ask Poppy chatbot, and a few other nifty things that smooth out daily operations. But we didn't have an employee-first app. We have this thick employee handbook packed with all the information employees need to know, but when someone has a question like 'How does accrued holiday time work?' they either have to ask Andrea or dig through pages of documentation. Poppy RAG on Slack solved this problem. Since we already use Slack for office communication, I built it as a Slack app that makes onboarding new employees so much easier and gives everyone instant answers to their exact questions without endless searching. The system uses RAG to query the employee handbook and retrieve relevant information, and I built in a web search tool in case they need information on California employment laws or workplace regulations. It's used pretty much every day, and the team loves how quick and easy it is to get answers right where they're already working.

Built for my wife Andrea's pediatric dentistry after seeing her overwhelmed during on-call weekends. Most calls weren't emergencies—just worried parents seeking reassurance. I created this chatbot as the first line of defense, trained on 750+ Q&A pairs verified by two other pediatric dentists. Uses a fine-tuned GPT-4o-mini model integrated via custom Voiceflow functions, including GuardGroq (LLaMA Guard-powered guardrails) and KnowledgeFlow (contextual retrieval inspired by Anthropic's approach). Presented at a Voiceflow developer meetup, KnowledgeFlow became popular for improving RAG accuracy. Result: Andrea now only gets emergency calls on weekends, making on-call shifts manageable. I am revamping this with a much more MCP forward version where patients can not only chat by text but also by voice as well as get details of their appointment and change their schedules or book a new appointment.

Built out of frustration with uploading private photos to remote sites for compression, POPMedia is a privacy-first media optimization tool where videos never leave your browser. Using FFmpeg.wasm running as WebAssembly, all video compression happens entirely client-side with no server uploads or cloud processing. Images are optimized server-side using Sharp for lightning-fast WebP conversion with up to 80% size reduction. The project served as a testing ground for Vercel as the go-to deployment platform for modern frontend applications, demonstrating how to build fast, privacy-focused web tools. Achieves 50-90% video file size reduction while maintaining quality, with batch processing support for multiple files simultaneously.

Contributed a comprehensive fix to browser-use that resolves duplicate local download events by ensuring each download is dispatched once. The solution hardens download handling across CDP progress, JavaScript fetch, and polling paths. Added a 'handled' guard to deduplicate dispatch events, disabled JS fetch fallback for local downloads to prevent double-dispatch, and improved remote handling by emitting completion using suggestedFilename when filePath is missing. Moved CDP handlers outside try blocks, implemented handler task tracking, and added proper cleanup on completion for enhanced stability.

Built for Poppy Kids Pediatric Dentistry, Roast My Snack reframes dental health as vanity rather than lectures—because teens care more about yellow teeth than cavities. Upload any snack photo and get a comic in ~30 seconds featuring Dr. Hawley, a molar tooth with Adult Swim energy who roasts snacks (never kids). The app uses Gemini 2.5 Flash for vision detection, Gemini 2.5 Pro with Thinking Mode for script generation, and Nano-Banana for comic image generation. Features include age-adaptive roast modes (Spicy for ages 9-12, Savage for 13-17), transparent dental risk scoring, taste-matched healthier swaps, allergen filtering, and kid-safe guardrails with 50+ blocked terms. Second place overall at the Sketch & Search hackathon.

KnowledgeFlow is a popular standalone Streamlit application that makes it easy to convert documents and tables for better RAG accuracy in Voiceflow. Inspired by Anthropic's contextual retrieval approach, it automatically processes and optimizes documents before uploading them to knowledge bases. Presented at a Voiceflow developer meetup, this tool has become widely adopted by the Voiceflow community. Notably, the innovative edit-before-upload feature—allowing users to review and modify documents before ingestion—was later adopted as an official Voiceflow feature. The app handles document chunking, metadata extraction, and contextual enhancement to ensure chatbots retrieve the most relevant information.

Built specifically for use with OpenAI's AgentBuilder platform to address limitations in the existing Outlook MCP server. The original AgentBuilder Outlook connector was read-only and only supported Graph authentication—not suitable for production environments. For the IntentIQ project, write access was essential for automated email responses and workflow routing. This comprehensive MCP server provides full read & write operations with dual authentication support: both Microsoft Graph API and organizational account credentials. Supports personal Microsoft accounts (@outlook.com) and organizational accounts (Microsoft 365) with multi-tenant architecture where users provide their own credentials. Handles token acquisition, payload validation, and complete email operations including send, reply, forward, delete, and folder management. Deployed on FastMCP Cloud for immediate use by MCP-compatible clients like Claude Desktop and Cursor. Features delegated permissions for personal accounts via Graph Explorer tokens and client credentials flow for organizational accounts with Azure AD app registration.