Damián Gil González - AI Projects & Products Portfolio
A curated collection of generative AI solutions — from enterprise multi-agent systems to real-time voice interfaces — built with cutting-edge models, frameworks, and cloud infrastructure.
Overview
10 Generative AI Projects Across Industries
From agriculture to enterprise analytics, these projects span the full spectrum of modern generative AI — multi-agent orchestration, RAG pipelines, synthetic data generation, voice interfaces, and beyond. Each was built with production-grade rigor, often before similar solutions existed in the market.
10
AI Projects
End-to-end generative AI products
3
Cloud Platforms
Azure, GCP, and more
6+
LLM Providers
Anthropic, OpenAI, NVIDIA, HF
4x
Speed Gains
Days of work reduced to minutes
Project 01
NL → Elasticsearch Agent
Built before similar tools existed in the market, this agent converts natural language queries into Elasticsearch DSL (Domain Specific Language) — enabling non-technical users to search complex datasets using plain language.
Key Innovation
A semantic correction layer validated key filter values before query execution, catching approximate or misspelled values that would otherwise return empty or incorrect results — a critical reliability improvement in production environments.
Stack
- LangChain / LangGraph orchestration
- Anthropic Claude as reasoning model
- Custom fuzzy value correction pipeline
- Elasticsearch DSL query generation
Project 02
GAIA — Multi-Agent System for Agriculture
GAIA is a WhatsApp-native multi-agent product designed specifically for the agricultural sector. Farmers and agronomists interact via WhatsApp to get instant answers on chemical products, pest control, crop management, and planting schedules — all powered by a fleet of specialized AI agents working in parallel.
WhatsApp Interface
Designed for accessibility — farmers interact in their native messaging app without any onboarding friction.
6–7 Parallel Agents
Agents specializing in chemicals, pests, crops, and plantations run concurrently to deliver comprehensive answers.
RAG + Fuzzy Matching
Vector database retrieval combined with string similarity correction handles orthographic errors common in field usage.
Projects 03 & 04
Enterprise AI at AXA — Synthetic Data & PDF RAG
Two complementary projects developed within the AXA corporate environment, deployed on Microsoft Azure infrastructure, addressing both data generation and document intelligence at enterprise scale.
🧪 Synthetic Dataset Generation
Created high-quality synthetic datasets to test and benchmark RAG systems in a controlled corporate setting. Leveraged NVIDIA Nemotron and specialized encoder models with a focus on code-domain data. This allowed safe evaluation without exposing sensitive real data.
- NVIDIA Nemotron + encoder models
- Code model specialization
- Microsoft Azure deployment
📄 Massive PDF RAG System
End-to-end development of a document ingestion and retrieval system capable of processing large volumes of corporate PDFs into a versioned vector database. Designed for reliability and traceability in regulated enterprise environments.
- Bulk PDF ingestion pipeline
- Version-controlled vector storage
- Full Azure cloud infrastructure
Project 05
Academic RAG — Universitat Politècnica de València
An end-to-end RAG system built for the UPV/UNICA academic community, indexing over 20 years of academic papers and research publications. Researchers and students can query the system in natural language to discover relevant articles, authors, and topics — dramatically reducing literature review time.
Document Processing
Ingestion, chunking, and vectorization of two decades of academic literature into a structured vector database for semantic retrieval.
Semantic Search Interface
Custom query interface allowing topic-based search and retrieval of academic papers, built entirely end-to-end including the front-end layer.
Full End-to-End Ownership
From raw document ingestion to user-facing interface — complete ownership of every layer of the architecture.
Projects 06 & 07
Deep Agent & MSP Architecture — Enterprise Orchestration
Two enterprise-grade multi-agent architectures designed for high-complexity environments, combining orchestrator patterns, NL-to-SQL, OCR pipelines, and distributed ticket processing.
🧠 Orchestrator Agent
Central agent delegating tasks to ~5 specialized subagents, each handling a distinct domain or data source within Azure infrastructure.
🗄️ NL to SQL
Natural language to SQL translation via Azure Databricks, enabling business users to query structured databases without writing code.
📷 OCR Pipeline
Document OCR pipeline extracting structured data and inserting it into Databricks databases — fully automated document digitization.
🎫 MSP Ticket Routing
Multi-agent service pipeline distributing incoming tickets across specialized agents based on content type, priority, and domain expertise.
Project 08 — Personal Project
AURORA — Multimodal RAG for Education
AURORA is a personal product in active development, pushing the boundaries of multimodal retrieval by processing video content from classes, master's programs, and professional training. Students can ask questions and receive contextually grounded answers drawn from hours of recorded lectures.
Multimodal Pipeline
- Pixeltable for video ingestion and frame extraction
- CLIP for joint text + image embeddings
- Qdrant vector database for storage and retrieval
- Streaming UI for real-time response delivery
- MCP (Model Context Protocol) as the central tool execution layer — the sole interface with the vector database, handling all search, indexing, and retrieval operations
Infrastructure
- ETL pipeline on Google Cloud Platform
- Scalable indexing of video, audio, and transcript data
- Streaming-first interface architecture
- Designed for scale across educational institutions
Projects 09 & 10
PDF Offer Analysis Pipeline & Real-Time Voice Agent
📑 Automated Tender Analysis
A multi-agent pipeline capable of simultaneously analyzing 100–300 PDFs — tender requirements alongside supplier offers with attached images and documents. The system automatically evaluates, ranks, and justifies scores for each provider.
A task requiring 4 days of human work is completed in approximately 10 minutes.
- Parallel multi-agent document processing
- Automated scoring with detailed justification
- Integrated UI for result visualization and review
🎙️ Real-Time Voice Interface
A real-time voice interaction layer for AI agents, enabling fully conversational experiences. Agents query vector databases live during the conversation to ground their responses in factual, retrieved data.
- Built with FastRTC (Hugging Face real-time framework)
- Full streaming mode for low-latency responses
- Vector DB integration during live audio sessions
- Python-native implementation
Tech Stack
General Technology Stack
Every project in this portfolio is built on a consistent, production-proven foundation of agent frameworks, frontier language models, vector databases, and cloud infrastructure — ensuring reliability, scalability, and state-of-the-art performance.
Agent Frameworks
- LangChain
- LangGraph
- LangSmith
LLM Models
- Claude (Anthropic)
- GPT (OpenAI)
- Nemotron (NVIDIA)
- Hugging Face models
Vector Databases
- Qdrant
- Pinecone
Cloud Infrastructure
- Microsoft Azure
- Google Cloud Platform
Interfaces
- WhatsApp
- Web streaming UI
- Real-time voice (FastRTC)