Álvaro – LLM, LangChain, Python

Large-scale enterprise platform for a US client in the bankruptcy and restructuring domain. The work centres on AI-assisted engineering, test automation, observability and release reliability for a four-person automation team. The team moved from manual, sequential test writing to an agent-driven pipeline that plans, generates and self-heals end-to-end tests.

• Builds and runs a multi-agent test generation pipeline: an orchestrator coordinates planner, generator and healer sub-agents working through the Playwright CLI.
• Assigns models by task cost: a lighter model handles mechanical page traversal, and the strongest model is reserved for healing tests that the generator failed to fix.
• Has shipped ~500 Playwright end-to-end tests in five months with this AI-assisted workflow.
• Builds internal AI tooling for the team, including a pull-request automation command and an AI code reviewer.
• Set up observability from scratch with OpenTelemetry, Grafana, Loki and structured logging. Nightly pipeline results are compared run-over-run to separate broken selectors from real regressions.
• Builds and improves CI/CD pipelines covering automated build, test and deployment.
• Develops and maintains Java and JavaScript services and APIs on the platform.

Self-serve platform that lets businesses build and deploy their own AI agents on WhatsApp without writing code. Users configure an agent through a web interface, connect their WhatsApp and OpenAI credentials, and the platform handles conversations, tool use and knowledge retrieval. Each agent can answer from the business's own documents and act on external services and databases.

• Owned core parts of the platform end to end in a lean team: requirements, technical decisions, deployment and post-release support.
• Designed the database architecture for agent configurations, parameters, conversations and per-user WhatsApp history.
• Built the no-code setup flow: users connected WhatsApp and OpenAI API credentials and adjusted agent behaviour without any code changes.
• Integrated the agent runtime with the Meta WhatsApp API, including conversation management and per-user context.
• Designed a tool layer with structured output schemas, so agents could safely read and modify data in third-party services and business databases.
• Built the RAG pipeline for agent knowledge bases: document processing, embeddings, vector search and contextual retrieval.
• When requirements expanded mid-project, separated orchestration from agent logic and reprioritised the highest-value changes, avoiding a full rewrite.
• Built backend services and workflow automations in Python connecting the platform to Supabase, Neon and n8n.

Online learning platform used by thousands of students, with course enrolment, lesson content and AI-powered learning support. Key features include an AI tutor chatbot that guides students through exercises without giving away answers. Instructor-side tools generate questions and lesson content from course material.

• Built the AI tutor chatbot with a RAG pipeline (LangChain + Chroma). Lesson content was stored in PostgreSQL and a vector DB, with retrieval scoped by lesson ID on lesson pages and hybrid search elsewhere.
• Improved retrieval quality by adding metadata, testing embedding models, adding reranking and building evaluation datasets from real student queries.
• Kept authorization at the session level: identity came from the user's session and API, never from model-supplied IDs, so the chatbot couldn't act on other students' data.
• Built AI features that generate quiz questions and lesson content from instructor input, using OpenAI and Gemini.
• Added in-platform prompt management so the team could adjust AI behaviour without code changes.
• Developed REST APIs for enrolment and student workflows, and resolved production bugs in enrolment and course functionality.

Public-facing conversational assistant that makes federal court case information accessible to citizens without navigating complex judicial systems. Users can query case status and procedural information by text or voice.

• Built the chatbot backend with LangChain and OpenAI to interpret user questions and generate contextual answers about court cases.
• Integrated judicial third-party APIs to retrieve and process live case data.
• Added speech-to-text and text-to-speech so users could interact with the assistant by voice.
• Built data-processing workflows connecting the conversational layer with external judicial systems.

Platform that digitizes forensic and medical examination workflows for the judicial system. It replaces complex paper forms with dynamic, structured digital forms that experts use to manage examinations end to end.

• Built backend microservices in Java and Spring Boot for examination management.
• Implemented dynamic digital forms that replaced manual examination paperwork.
• Integrated Java microservices with Python-based AI agents to support conversational interfaces and automated steps.
• Designed REST APIs and database integrations for core examination workflows.
• Containerized services with Docker and supported integration across the platform.