Onde:
AI-Powered Real Estate Platform
A full-stack real estate marketplace featuring advanced AI search capabilities, semantic property discovery, and intelligent chat assistance. Built with a modern dual-architecture approach combining traditional REST APIs with serverless edge functions.
OpenAI Integration
► GPT-4o-mini for conversational AI chat and intent classification
► GPT-4o Vision API for automated property image analysis, extracting detailed visual descriptions of interiors, exteriors, colors, materials, and architectural features
► text-embedding-3-small (1536 dimensions) for property text embeddings
► text-embedding-3-large (3072 dimensions) for image and document embeddings
► Intelligent filter extraction from natural language queries
► Context-aware property recommendations
Semantic Search Pipeline
► Multi-stage AI processing pipeline with <3s response time
► Parallel intent classification and filter extraction
► Visual search mode for queries like "houses with green floors" or "properties with gardens"
► POI (Point of Interest) proximity search integration
► Smart context building from vector matches
Structured Data Output
► Zod schema validation for API responses
► Type-safe filter extraction with explicit TypeScript interfaces
► Structured JSON responses from AI models using function calling
► Deterministic property reference detection ("la primera", "that one")
Pinecone Vector Database
► Three specialized indexes for different content types:
► Properties Index (text-embedding-3-small, 1536 dim): Property descriptions, addresses, features
► Images Index (text-embedding-3-large, 3072 dim): Visual analysis from property photos
► Documents Index (text-embedding-3-large, 3072 dim): Floor plans, contracts, certificates
► Parallel vector search across all indexes
► Metadata filtering (price, bedrooms, location, amenities)
► Deduplication by property_id
► Hash-based vector IDs for idempotent operations
Embedding Generation System
► Automated batch processing (5 images parallel, up to 50 images per property)
► Deterministic SHA-256 hash-based vector IDs to prevent duplicates
► Image analysis with detailed prompts for exterior/interior features
► Property text chunking with semantic sentence splitting
► Fire-and-forget async embedding generation with error tracking
Main AI chat orchestrator
► 5-step processing pipeline optimized for performance
► Parallel intent classification and filter extraction
► Vector search coordination across Pinecone indexes
► GPT-4o-mini response generation with context awareness
► Conversation history management with session tracking
Embedding generation service
► OpenAI Vision API integration for image analysis
► Multi-index embedding distribution
► Automatic metadata normalization
► Chunk-based text processing
► Document storage in Supabase with preview generation
Natural language query parser
► Intent classification (search, POI, compare, filter_update)
► Structured filter extraction from natural language
► AI filter tagging (pet-friendly, modern-kitchen, quiet-neighborhood)
► Region lookup with case-insensitive database matching
► Property reference detection
Semantic search coordinator
► Parallel embedding generation (small + large models)
► Multi-index vector search
► PostgreSQL property fetching with region filtering
► Result deduplication and scoring
► Google Places API integration for POI queries
Google Maps API
► Places API for POI proximity search
► Geocoding for address validation
► Distance matrix calculations
► Walking time estimations
► Place details with ratings and reviews
Dual Architecture System
Express.js API (Node.js)
► REST endpoints for property CRUD operations
► JWT + Supabase Auth integration
► Socket.io real-time chat with Redis pub/sub
► Worker clustering with per-worker connection pools
► CSRF protection and security middleware
Deno Edge Functions (Supabase)
► Serverless AI chat and semantic search
► Low-latency edge deployment
► Multi-tier caching strategy (in-memory FIFO)
► Automatic scaling and cold start optimization
Database & Storage
► PostgreSQL with PostGIS for geospatial queries
► Kysely type-safe query builder
► Connection pooler for production scaling
► Supabase Storage for property images
► Redis for caching and pub/sub
Performance Optimizations
► Multi-tier caching (embeddings: 24h, properties: 10min, searches: 5min)
► Batch parallel processing for image embeddings
► Debounced search inputs (500ms)
► Query result prefetching on hover
► localStorage persistence for chat history
React 19 + TypeScript + Vite
► Modern React patterns with hooks
► Type-safe development with strict TypeScript
► Fast HMR with Vite
State Management
► TanStack Query v5 for server state
► Optimistic updates for favorites and offers
► Smart caching with staleTime configuration
► URL params as single source of truth for search state
UI & Styling
► Tailwind CSS for utility-first styling
► ShadcnUI + Radix UI component library
► Custom design system (Montserrat, Saira, Inter fonts)
► Responsive design with mobile-first approach
Key Features
► AI-powered conversational search with persistent chat history
► Real-time property chat with Socket.io
► Interactive property cards with clickable AI chat links
► Region-based and radius-based search
► Advanced filters modal with amenities and conditions
► Favorite properties with optimistic updates
► Offer negotiation system
► Visit scheduling
Type Safety
► End-to-end TypeScript coverage
► Generated database types with Kysely
► Zod runtime validation
► Strict ESLint + Prettier configuration
Build & Deploy
► Path aliases for clean imports (@/, @controllers/, @services/)
► Environment-based configuration
► Hot reload in development
► Production clustering with graceful shutdown
Testing & Monitoring
► Detailed console logging in edge functions
► Winston structured logging in Express
► Performance tracking (intent classification, cache hit rates)
► Error boundary handling
► AI Response Time: <3 seconds end-to-end
► Embedding Processing: 5-10 minutes for 50 images (batch parallel)
► Vector Search: <500ms across 3 Pinecone indexes
► Cache Hit Rate: ~70% reduction in repeated queries
► Database Queries: Type-safe with <100ms response time
► Real-time Chat: Sub-second message delivery with Socket.io
1. AI Chat with Memory - Persistent chat history using localStorage, property reference context
2. Semantic Visual Search - "Show me properties with marble kitchens" using image embeddings
3. Natural Language Regions - "Properties in Avenidas Novas" auto-completes region selection
4. Smart Filter Detection - Extracts price, bedrooms, amenities from conversational queries
5. Clickable Property Links in Chat - Navigate to property details directly from AI responses
6. Batch Parallel Embedding - Process 50 images in 5-10 minutes with deduplication
7. Multi-Index Vector Search - Combines text, image, and document embeddings for comprehensive results
8. POI Proximity Intelligence - "Properties near hospitals" with walking distance calculations
SugarSync:
AI-powered diabetes management ecosystem focused on personalized nutrition, glucose insights, and doctor–patient collaboration.
Cross-platform app
► Built with React Native, Expo and TypeScript for iOS and Android
► Flows for daily logging of meals, glucose readings and physical activity
► Mobile-first UX for quick entry, reminders and progress tracking
Patient & doctor views
► Separate experiences for patients and clinicians
► Clinician dashboards to review trends, notes and alerts
► Secure sharing of records for remote follow-up
Food & nutrition intelligence
► Integration with USDA and Edamam APIs for macronutrients and carb counting
► Voice- and text-based food input for faster logging
► Normalization of ingredients and portions into structured nutritional data
Predictive glucose insights
► Pinecone + OpenAI embeddings to correlate meals, activity and glucose history
► RAG-style retrieval of similar days/meals to suggest risk or expected response
► Smart suggestions for meal timing and composition based on historical patterns
Supabase platform
► PostgreSQL schema for users, clinicians, logs, alerts and recommendations
► Row Level Security (RLS) for strict multi-tenant data isolation
► Supabase Auth for account management and session handling
Edge Functions for AI
► Supabase Edge Functions (Deno) to orchestrate OpenAI calls and Pinecone queries
► Serverless pipelines for nutritional analysis and glucose insight generation
► Structured JSON outputs validated with Zod for type-safe APIs
Security & compliance mindset
► Data flows modeled following HIPAA-like principles (least privilege, auditability)
► Separation of PHI-like data from analytics metadata
► Encrypted storage for sensitive information where applicable
1. AI-assisted carb counting via USDA/Edamam with voice input
2. Doctor–patient shared workspace with historical context and notes
3. Predictive insight cards for "today's risk" based on past days
4. Event-driven recommendations running on serverless Edge Functions
5. RLS-protected multi-tenant architecture ready for clinical pilots
Spark n Co:
In this project, I contributed to integrating an AI-powered chat solution using OpenAI and DeepSeek to automate and streamline Jira ticket management.
The system leverages advanced natural language processing to interpret user inquiries in real time, automatically generating tickets, creating incident reports, and dispatching notifications without manual intervention.
Pursuit:
I led the development of a mobile app using React Native and Expo, which integrates Supabase for backend services and authentication.
I also implemented a Retrieval and Augmented Generation (RAG) model with Pinecone to store personalized training and nutrition plans based on each user's personal data, all generated with an OpenAI AI model.
The app combines AI, structured data output, and data analytics to provide personalized recommendations that optimize individual performance and overall well-being.
In the process, I trained the team on:
► Generative AI model APIs, cost-based token counting, context injection, prompting structure management, and structured data output.
► Created embeddings and similarity functions to generate similarity recommendations and searches.
► Created Supabase edge functions using Deno.js and optimized them to interact with OpenAI request times.
Rentscape:
I developed a feature within the Rentscape platform that features an AI-powered chat solution.
This system leverages chat interaction with AI combined with function_calling and structured data output to generate accurate data as the AI obtains data with each message.
That data is used to generate embeddings and perform Pinecone searches for the best rental properties.
The solution offers an intuitive and personalized search experience based on a mathematical calculation, streamlining the decision-making process in the real estate market.
In the process I instructed the team on the following:
► Use of AI reasoning models that allow function_calling and structured data output at the same time.
► Vector database management and client creation for Pinecone using Supabase's Edge functions with Deno
The responsibilities are: Pre-select new Teach Assistants. Train them in soft skills. Conduct role-plays, where the pre-selected candidates participate in a simulated situation. Finally choose the finalists, who will be in charge of the incoming students to Henry's bootcamp.
Develop a dynamic page using AWS (DynamoDB, S3, Lambda), NextJs, Zustand, Tailwind, MaterialTailwind Integrate rich text markdown system and multimedia file management. Authenticate with token and controlled forms. Develop with SCRUM methodologies, exercise GitFlow, coordinate with Notion task board. Optimize site functionality, achieve agile and dynamic user interface, achieve attractive aesthetics according to home page standards and representative of a company that provides technology services. Propose creative ideas and solutions for process improvement.
Main tasks: Coordinate a group of new joining students throughout the whole bootcamp experience. Guide students in the first steps of the course and clarify their doubts. Provide assistance in solving exercises and encourage group teamwork (Pair Programming). Suggest ideas to improve the Bootcamp processes. Promote collaboration and build community within the students.
In this position, I started a fullstack project from scratch, as part of my own bootcamp project, using React, Redux, Html and Css for the frontend; node.js and express with morgan as middleware for the backend, with the challenge of a faulty API, with frequent crashes, and updates that changed the structure, here I learned how to create my own API on github to make support requests, and structure the code for proper error handling and to work whether the api is down or not.
I started by integrating the knowledge of the bootcamp in the frontend such as React, Redux, Html and Css and in the backend with express in a bootcamp integrator project, which was modified in various ways throughout the bootcamp itself, advancing in the functional part, and in the cards it is possible to see the visual progress.
I programmed, updated and repaired 3D printers FDM
A full-stack real estate marketplace featuring advanced AI search capabilities, semantic property discovery, and intelligent chat assistance. Built with a modern dual-architecture approach combining traditional REST APIs with serverless edge functions.
A cross-platform, AI-powered diabetes management app focused on personalized nutrition, glucose insights, and doctor–patient collaboration. Built with React Native, Expo and Supabase, it integrates USDA and Edamam APIs for carb counting, voice-based food logging, and structured nutritional analysis. Using OpenAI and Pinecone, SugarSync correlates meals, activity and historical glucose data to surface predictive insights and "today's risk" cards, all on top of a secure, RLS-protected multi-tenant backend ready for clinical use.
A mobile application developed with React Native and Expo, integrating Supabase for backend services and authentication, and utilizing OpenAI. It implements a Retrieval Augmented Generation (RAG) model with Pinecone to generate personalized training and nutritional plans based on each user's personal data. The application combines AI and data analysis to provide tailored recommendations that optimize individual performance and overall well-being.
An integration within the Rentscape platform that features an AI-powered chat solution. This system leverages interactive chat combined with function calling to generate precise data, which is then used to perform searches on Pinecone for the best rental properties. The solution offers an intuitive, personalized search experience, streamlining the decision-making process in the real estate market.
Dynamic blog delivered for incorporation to the company website. Collaboration between students applying the "scrum" methodology. User Interface built on Next.js 13, with Zustand implementation for global state management. Server built on AWS Lambda, configuring AWS Api Gateway as event trigger. Image storage system on AWS S3 Buckets.
App developed with React Native and Expo using the Pokemon api, to display information about pokemon. Data persistence is handled with the AsyncStorage library.
I developed an app with React Native CLI and TypeScript to view the pictures taken by the Hubble telescope provided by NASA's 'Astronomy Picture of the Day' api.
App developed with React Native Expo and Tailwind is a simple but effective tool to improve productivity. With a minimalist and easy to use interface, this application allows users to divide their time into work and rest periods using the Pomodoro technique.
My name is <Joaquin Emanuel Lizarraga />, but my friends call me Ema. I wrote my first program on a PIC using a PIC burner in C when I was 14 years old. After that, I moved on to Arduino, building home automation systems and robotics with IoT. Now, I create applications that interact with sensors and leverage AI.
Currently, I'm developing SugarSync, an app for diabetes control and management that enables users to read data from various sensor brands, track their carbohydrate intake, and maintain a comprehensive record accessible to their doctors.
Among my proudest achievements is my experience during Henry's bootcamp, where I was elected as a Teaching Assistant. I guided and supported new bootcamp entrants through an intensive path to graduation, achieving a 100% success rate. Today, I am a developer specializing in AI implementations for mobile applications, working on innovative projects that integrate AI-powered chat solutions, automated ticket management, and other cutting-edge technologies.
