Alexis – AWS, Python, LLM

An online platform that helps create on-brand emails for every stage of client–customer communication, from activation, to engagement and retention. It automatically scans the user website to extract its branding elements, then uses AI to generate pertininent copy.

• Translated the product vision into a working platform from the ground up (apart from graphic design).
• Extracted branding assets (colors, fonts, images, etc.) from customer websites for use in automated emails.
• Converted existing email examples into modular templates, customizable with brand elements and AI-generated copy.
• Applied prompt engineering to generate tailored copy for ~300 lifecycle communications, aligned with specific templates.
• Set up email delivery systems and managed backend services with FastAPI in Docker on Google Cloud.
• Built the infrastructure from scratch using Terraform.
• Designed and implemented data models in Supabase (users, organizations, billing, etc.).

On the front-end side:

• Developed an email editor for further customization of the generated communication.
• Integrated Stripe to manage credits and subscriptions.
• Connected the user dashboard with backend services via Supabase.

The product is an autonomous Customer Support Agent designed for expense management in businesses (B2B). It assists users with tasks such as retrieving information about user approvals, creating expense rules, and analyzing data.

• Designed the worflow to respond to the different use cases: data retrieval, analysis, and configurations;
• Implemented the system from scratch using LangGraph;
• Managed all prompts necessary, tailoring the information provided for each step (explanation of the business, relevant data schemas, etc.) while measuring the effectiveness of nodes separately;
• Developed an intermediate backend to be provided as tools for the agent. This backend would interact, in turn, with the main company backend;
• Deployed the backend on ec2 using Docker.

Improved the client's internal AI usage by implementing a Retrieval Augmented Generation system that integrates the company's existing knowledge about customers and product development from their knowledge base, CRM, and internal database.

• Set up conversion of documents to embeddings using OpenAI's embeddings API;
• Stored vectors in superbase and set up vector similarity functionality;
• Pipelining for continuous integration of updated Knowledge Base and CRM;
• Built search with entity recognition to convert queries into embeddings for lookups;
• Lowered costs by storing all embeddings already obtained;
• Offered via API with options to choose the source of the augmented context: KB, CRM, or Data Warehouse.

The product is a platform for knowledge enthusiasts to discover and create quizzes. AI assists by helping users choose topics, set difficulty levels, and adjust question types, while also identifying engaging titles and hashtags. Its recommendation model personalizes quiz suggestions based on user preferences and promotes new content.

• Developed the recommending system using Pytorch's Torchrec and exposed it through an API (in the style of "Tiktok's for you feed");
• Developed the Data pipelining to feed the recommender in real time;
• Set up and implemented the agentic workflow to assist quiz creation;
• Designed tests to ensure the factuality and diversity of our quizzes;
• Deployed everything on EC2, using FastAPI, Docker, and CI/CD;
• Managed all the infrastructure with Terraform;
• Implemented semantic search to find quizzes based on keywords;
• Ensured security and scalability of the backend, managing authentication, load balancing, and alerts.

The client is a startup that buys invoices from Chilean companies, which the recipients are to pay. Historically, invoice reviews have experienced a high rejection rate based on various risk factors. The project's goal was to develop a machine-learning model to evaluate these assets and generate risk-based proposals.

• Designed the proposed solution;
• Defined success metrics, identified risks, and aligned stakeholders;
• Led the development, assisted by a more junior data scientist;
• Developed data pipelines to feed the model using data;
• Deployed the solution on GCP to be called directly from the product.

Minute of glory:

• The machine learning model approach reduced the rejection rate by 10% in the first version and another 10% in the second iteration.