Federico – AI agent orchestration, Python, LLM

Enterprise HR teams in LATAM spent most of their time on repetitive work: job descriptions, skill assessments and career planning. Federico architected a multi-tenant SaaS platform where specialised AI agents automate up to 70% of daily HR operations for thousands of users, freeing HR teams for strategic work.

The system delivers personalized employee career development plans using multi-agent AI, helps organizations transition to skills-based models by identifying competency gaps, and automates job description creation and updates. It also supports AI-driven skill assessments for workforce evaluation and streamlines repetitive HR workflows, enabling HR teams to focus on higher-impact strategic work. The platform is built as a multi-tenant SaaS solution with enterprise-grade security, compliance controls, and real-time integrations with existing HRIS systems.

• Designed end-to-end AI solution architecture from data ingestion through deployment and operations, integrating multi-agent ecosystems, RAG pipelines, and LLM orchestration into a unified enterprise platform;
• Architected a multi-agent system using LangGraph with tool-calling agents handling complex cross-domain HR workflows: career path generation, skill gap analysis, job description automation, and employee performance evaluation;
• Built and optimized RAG pipelines with embeddings and vector search (Pinecone, Weaviate) for knowledge grounding across HR policy documents, competency frameworks, and organizational data — reducing hallucinations by 40%;
• Optimized inference costs by 60% through strategic prompt caching, batch processing, and model tiering strategies across Claude and GPT-4 deployments;
• Deployed cloud-native infrastructure on AWS (Lambda, ECS, S3, RDS) with containerization, secrets management, and scalability planning, achieving 99.9% uptime;
• Implemented a full observability stack (LangSmith) with production dashboards tracking cost per tenant, model performance, retrieval quality, and pipeline health;
• Established MLOps/LLMOps practices with CI/CD pipelines (Terraform, GitHub Actions), automated testing, model versioning, and drift detection;
• Embedded AI governance framework with human-in-the-loop controls and GDPR-aligned data handling for regulated HR environments;
• Built and grew the engineering and AI team from zero, defining technical standards, architectural patterns, and delivery processes.

A Mexican cleaning-products brand handled product questions, complaints and promotions by hand. Federico built one WhatsApp agent covering support, product recommendations and outbound campaigns, reducing support load for routine purchase decisions and giving the operations team real usage data.

The system features:

• FAQs and complaints handling.
• Product recommendation engine that interprets natural language descriptions of cleaning needs (e.g., stain type, fabric color, floor or surface type) and suggests the most suitable products along with usage instructions.
• Outbound promotional campaigns.
• Quality complaint intake.
• Escalations to human agents when necessary.
• An analytics dashboard providing the operations team with insights into customer behavior and product performance.

• Architected a conversational AI agent on WhatsApp Business API handling three distinct flows: inbound customer support (FAQ and complaint management), intelligent product recommendation, and outbound promotional messaging campaigns;
• Built a product recommendation engine using an RAG pipeline over the product catalog, matching user-described needs (stain type, surface material, fabric color) to the appropriate SKU with usage guidance — reducing support load for routine purchase decisions;
• Designed quality complaint intake workflow with structured data extraction, automatic severity classification, and escalation routing to human agents for unresolved cases;
• Implemented outbound promotional messaging system allowing the marketing team to trigger segmented campaigns through the same WhatsApp channel used for support, maintaining consistent brand voice via LLM-generated personalization;
• Built and maintained Pinecone-backed knowledge base covering product catalog, usage guides, FAQ content, and complaint resolution playbooks;
• Deployed production FastAPI backend on GCP with async session management via LangGraph to handle concurrent conversations without state collision;
• Built a React and TypeScript analytics dashboard surfacing customer interaction trends, most-queried products, complaint categories, recommendation acceptance rates, and campaign engagement - enabling the team to refine both product positioning and support processes from real usage data.

Mexico's federal intellectual property authority needed small shop owners and consumers to be able to check whether a product is legitimate without specialist knowledge (a focus on the tobacco sector). Federico built a WhatsApp-native agent that takes photos, documents and questions and returns an instant legality assessment, making enforcement accessible through a channel people already use.

Users can submit product photos for visual authenticity analysis, upload documents for validation, and ask natural language questions about intellectual property regulations and product legitimacy. The system combines a conversational AI support agent with a computer vision layer for real-time product inspection.

• Architected end-to-end conversational AI system integrated with the WhatsApp Business API, handling text, image, and document inputs within a single unified agent flow built on LangGraph;
• Designed a multi-modal agent pipeline that routes user inputs by modality: text queries to the RAG-powered FAQ and regulatory knowledge base, product photos to a vision model for legality classification, and documents to a parsing and validation workflow;
• Built and maintained a RAG knowledge base (Pinecone), ingesting IMPI regulatory content, product registration databases, and IP law documentation to power accurate, grounded responses on legal standing and compliance;
• Implemented product image analysis pipeline enabling users to photograph a product label or packaging and receive an automated legality assessment - targeting counterfeit tobacco and unlicensed goods detection
• Deployed production-grade FastAPI backend on GCP with asynchronous message handling to support concurrent WhatsApp sessions without latency degradation;
• Designed conversation state management with LangGraph to maintain multi-turn context across WhatsApp sessions, allowing users to iteratively refine queries and receive follow-up analysis without re-explaining context;
• Built document intake flow for users to submit invoices, permits, or product registrations and receive structured validation feedback against official IMPI records;
• Optimized system for non-technical end users (small shop owners, market vendors) through simplified UX flows, Spanish-language responses, and graceful fallback to human escalation when confidence was below threshold;
• Built a full analytics dashboard in React and TypeScript, giving IMPI teams visibility into platform usage patterns, query categories, product flagging rates, and counterfeit detection trends — enabling data-driven decisions to improve enforcement processes and identify high-risk product categories or geographic clusters.

A Saudi law firm struggled to keep pace with frequent legislative changes across several regulatory domains. Federico built an English/Arabic copilot that ingests new and amended laws automatically and supports contract review, drafting and case-law analysis, so lawyers always work from current law.

The system continuously monitors official legal sources, automatically ingests and processes new and amended legislation, and maintains a production-grade RAG knowledge base updated in near real time. Deployed as a domain-specific assistant for practicing lawyers, it supports contract review and risk analysis, automated contract drafting from structured templates, and in-depth analysis of past judicial rulings and case outcomes. The platform is accessible in both English and Arabic through a conversational interface.

• Designed and built an automated ingestion pipeline that tracked Saudi legislative sources and processed weekly law amendments across multiple regulatory domains, keeping the knowledge base continuously up to date without manual intervention;
• Engineered domain-specific chunking and parsing strategies for legal documents, respecting article, clause, and section hierarchies to preserve legal context and prevent cross-boundary information bleed;
• Built a production-grade hybrid RAG system combining dense (embedding) and sparse (BM25/keyword) retrieval with reranking, optimized for precise legal terminology retrieval in both Arabic and English;
• Implemented a multilingual embedding pipeline with Arabic-aware tokenization and right-to-left text handling, ensuring retrieval quality was equivalent across both languages;
• Developed contract review agent: structured clause extraction, risk flag identification, and comparison against jurisdiction-specific legal standards;
• Built a contract drafting assistant with template-aware generation and compliance validation against current Saudi regulatory requirements;
• Created a case law analysis module enabling semantic search over past judicial rulings, precedent identification, and outcome pattern analysis to support litigation strategy;
• Deployed production API with role-based access controls and full audit logging to meet the confidentiality and compliance requirements of a regulated legal environment.