Alex – Python, LLM, AI

At this company, marketing teams lacked the ability to measure the true causal impact of digital advertising campaigns across channels like Meta, Google Ads, and Amazon. Existing tools provided fragmented insights and often misattributed success.

Alex built a multi-agent causal measurement platform to unify cross-platform data, apply causal inference techniques, and provide actionable insights on incremental ROI. Specialized agents collected and normalized data, ran causal experiments (geo-lift, conversion lift), and analyzed treatment effects. Context-aware sentiment agents assessed ad placement safety and emotional alignment. An orchestration layer coordinated all agents, while validation loops ensured accuracy and robustness.

The system translated technical causal outputs into executive-friendly insights via natural language generation and dashboards, driving millions in efficiency gains. Results included +45% measurement accuracy over last-click attribution, $3.2M in waste reduction, and a 28% ROI lift. The system became the foundation for an annual $15M efficiency gain.

• Designed and implemented a multi-agent system for causal measurement across advertising platforms.
• Built specialized data collection agents for Meta, Google Ads, Amazon, and analytics platforms.
• Developed normalization pipeline for cross-platform metrics (CPM, CPC, ROAS).
• Implemented causal inference methods (Bayesian time-series, causal forests, DAGs, geo-lift tests).
• Built transformer-based agents for context & sentiment analysis of ad placements.
• Designed orchestration and decision engine with Apache Airflow and optimization models.
• Deployed system on AWS ECS with Kafka event streaming, Redis caching, and Snowflake data storage.
• Set up monitoring with Datadog dashboards and anomaly detection in causal estimates.
• Created natural language summaries and dashboards for executives with Plotly Dash.
• Achieved 92% accuracy in incremental conversion prediction and reduced decision latency from 2 weeks to 2 hours.
• Scaled the system to process 50M+ events daily, including TV attribution and retail media.

An internal AI chatbot for a large enterprise client. The chatbot needed to provide accurate answers about internal documentation and policies while responding in the distinct tone of company executives (e.g., CTO, Design Director) to build trust and engagement.

The solution combined Retrieval-Augmented Generation (RAG) with fine-tuned persona adapters. Alex designed a retrieval layer using semantic embeddings and Pinecone for real-time answers (<500ms latency), and fine-tuned LoRA adapters per persona to replicate communication styles. An evaluation framework ensured accuracy and tone alignment, while feedback loops allowed continuous improvement.

The chatbot significantly reduced HR and IT support tickets, increased internal NPS, and gained executive sponsorship for expansion into new teams and personas

• Designed and implemented the RAG pipeline with Pinecone and OpenAI embeddings.
• Built fine-tuned LoRA persona adapters using synthetic + real data.
• Integrated persona routing into LangChain’s generation module.
• Implemented evaluation pipeline (LLM-as-a-judge, accuracy benchmarks, latency tracking).
• Designed an active learning loop with user feedback and retraining.
• Ensured production safety with content moderation and prompt filters.
• Achieved 85%+ persona tone accuracy (vs. <60% baseline).
• Reduced support ticket volume by 40% and improved NPS by +25 points.

At this company, support teams struggled to navigate a large internal knowledge base due to keyword-based search limitations. Alex built a semantic search system powered by RAG and LLMs to provide context-aware answers using natural language queries.

The pipeline combined embeddings and Pinecone for retrieval, dynamic prompting via LangChain, and a reranker for higher precision. Guardrails detected hallucinations and unsafe outputs, with fallback to links or human escalation. The system was deployed on AWS (SageMaker + Lambda), integrated into the support agents’ UI, and included monitoring with human-in-the-loop review for continuous improvement.

The project improved retrieval precision by ~35%, reduced resolution times by 25%, and increased agent satisfaction by 40%. The same architecture was later scaled to HR and finance teams

• Collected and processed ~100K internal support articles, FAQs, and chat logs.
• Designed semantic chunking with recursive text splitting.
• Tuned embeddings on in-domain data using PEFT.
• Implemented RAG pipeline with Pinecone + LangChain.
• Added a reranker transformer for precision improvement.
• Built guardrails for hallucination and unsafe output detection.
• Deployed production system on AWS (SageMaker, Lambda, Step Functions).
• Designed monitoring with human-in-the-loop feedback and a retraining loop.
• Achieved +35% retrieval precision and -25% resolution time.
• Scaled architecture to multiple departments (HR, Finance).

At Brazil’s largest beauty retailer, Alex's work centered on designing and implementing high-performance Retrieval-Augmented Generation (RAG) systems that seamlessly integrated generative AI models, driving insights and enhancing data accessibility across the organization. Focused on maximizing value, Alex delivered robust data pipelines that ensured reliable, efficient, and impactful data-driven solutions for various business needs.

• Developed and optimized generative AI solutions, including intelligent chatbots and recommendation engines, leveraging LLMs and prompt engineering to improve response relevance and user engagement.
• Integrated vector databases (e.g., Pinecone, ElasticSearch) to enhance embedding storage and retrieval, driving increased accuracy and scalability in search and recommendation functionalities.
• Built and maintained robust data pipelines to support high-volume data retrieval across internal and external systems, ensuring seamless and low-latency information flow.
• Employed Docker and Kubernetes for containerized deployment of back-end services, establishing stable and resilient infrastructure across distributed environments.
• Collaborated with data science and engineering teams to create and deploy reusable libraries and frameworks, streamlining generative AI model integration with the back-end. Achievements:
• Successfully delivered 10+ high-impact generative AI projects, resulting in a 20% improvement in operational efficiency and a measurable increase in profitability.
• Optimized data pipeline architecture, reducing processing times by 25%, which enabled faster, more efficient real-time data access.

At Brazil’s largest beauty retailer, knowledge management was a major bottleneck across 4,000+ franchise stores and corporate teams. Employees wasted time searching for product information, policies, and procedures across fragmented systems, creating inefficiencies, duplicate tickets, and inconsistent customer service.

The solution was an Enterprise Knowledge Assistant integrated directly into Slack, the company’s main communication platform. This assistant retrieved accurate information from 50,000+ internal documents and handled 10,000+ daily queries in Portuguese with <2s response times.

Key features included:

• Multi-agent architecture with semantic retrieval, contextual enrichment, and continuous learning.
• Native Slack bot with slash commands, interactive responses, multilingual support, and proactive notifications.
• Scalable AWS deployment with ECS, Lambda, RDS, and Pinecone vector DB.
• Robust monitoring, analytics dashboard, and user feedback loops.

The project reduced response times from 45 minutes to under 2 seconds, cut duplicate tickets by 60%, improved sales conversion, and became the AI template for other business units.

• Designed and implemented a multi-agent architecture using LangGraph.
• Built a scalable knowledge ingestion pipeline for 50,000+ documents (PDFs, images, structured data).
• Integrated Slack bot with advanced features: commands, contextual responses, and multilingual switching.
• Deployed auto-scaling AWS infrastructure (ECS Fargate, Lambda, RDS, S3, CloudFront).
• Ensured full LGPD compliance and enterprise security standards.
• Created real-time analytics dashboards for usage, response quality, and business impact.
• Reduced response time from 45 min to <1.8s; achieved 92% accuracy in Portuguese queries.
• Drove adoption by 85% of employees within 3 months.

Alex worked as a Data Scientist specializing in demand forecasting and inventory management. His role involved analyzing data to uncover patterns and trends affecting product demand, collaborating with cross-functional teams to develop data-driven solutions, and overseeing the entire lifecycle of forecasting models. Alex provided actionable insights to optimize inventory management, reduce stockouts, and enhance operational efficiency.

• Conduct exploratory data analysis to identify patterns, trends, and relevant insights.
• Managed the complete lifecycle of forecasting models, from data collection and preparation to evaluation and deployment in production.
• Provide actionable insights and data-driven recommendations to optimize inventory management and reduce stockouts in stores.