Lou – LLM, LangChain, RAG

A shared, channel-agnostic AI platform that serves as the core intelligence layer across multiple communication channels and product domains. The system was designed to process text-based interactions independently of the delivery channel, enabling consistent conversational logic, context management, and business workflows while supporting reuse across different products and use cases.

• Unified existing channel-specific AI implementations into a shared, channel-agnostic core, enabling the same underlying engine to be reused consistently across multiple communication channels and product verticals;
• Led the design of a decoupled conversational intelligence architecture, separating AI logic from channel-specific concerns to improve scalability, maintainability, and reliability;
• Designed and implemented conversation scoring, evaluation frameworks, and quality metrics, making conversational performance measurable across AI-powered products;
• Worked at the intersection of AI architecture, engineering, and product strategy, helping define how conversational AI systems were built, evaluated, and scaled across the organization;
• Served as a technical reference for AI systems design, platform extensibility, and long-term architectural direction.

An AI-powered conversational platform that automated customer interactions for the residential leasing process via SMS. The system handled inquiries, guided prospective tenants through the rental journey, answered property-related questions, and supported lead qualification using large language models and conversational AI workflows.

• Led a from-scratch rewrite of the AI engine powering Zuma’s leasing assistant over SMS;
• Designed and implemented an agentic system composed of tool-calling agents, specialized sub-agents, and an internal knowledge base;
• Significantly improved conversational quality, reliability, and extensibility, enabling faster iteration and safer evolution of AI behavior;
• Increased conversation success and conversion rates by improving intent handling, reasoning flow, and response consistency;
• Established architectural foundations that later enabled reuse across channels and products.

An in-house RAG system that let users upload text-based documents and chat with them from pre-defined seed prompts.

• Built a proof of concept for an internal RAG system enabling users to upload text-based documents and interact with them through chat interfaces using predefined seed prompts;
• Implemented document metadata storage in MySQL to support indexing, retrieval, and system-level organization of uploaded content;
• Used Amazon S3 as the primary storage layer for document text data, ensuring scalable and durable object storage;
• Integrated Qdrant as a vector database for storing and querying embeddings to enable semantic retrieval functionality;
• Leveraged OpenAI models for both embedding generation and natural language response synthesis within the RAG pipeline;
• Developed a Python-based FastAPI backend exposing ingestion and retrieval endpoints for document processing and conversational querying.

A project health analytics and contributor intelligence product across 15,000+ open-source repos.

• Owned and delivered department-wide initiatives, providing regular progress updates to senior leadership;
• Designed and monitored detailed project schedules, assigning responsibilities across multiple teams;
• Led proposal development efforts, presenting technical opportunities for go/no-go decisions at the leadership level;
• Directed a cross-functional team of approximately 15 members, including engineering, QA, and DevOps, spread across the Americas to Asia-Pacific time zones;
• Collaborated closely with Product leadership to align on priorities, roadmap execution, and long-term strategy;
• Established efficient task assignment frameworks, reducing redundancy and minimizing interdependencies;
• Attended weekly department head meetings to communicate updates, surface risks, and ensure cross-team alignment;
• Worked directly with vendors and internal sourcing teams to explore new opportunities, optimize solutions, and identify cost-saving measures.

Led the distributed Engineering Team in building a product in the Software Engineering Intelligence space and pioneered the work on AI agents since late 2022.

• Led the distributed Engineering Team (9 persons).
• Defined the technological strategy of the business as part of the leadership team;
• Provided strong technical leadership for architectural decisions based on business needs;
• Was responsible for the career plan and the growth of the members of the Engineering Team;
• Created hiring plans of the Engineering Team;
• Was in charge of the Engineering spend.

Achievements:

• Successfully built a healthy and collaborative working environment by fostering clear communication, and knowledge sharing;
• Successfully contributed to the system architecture’s definition and planning thanks to the expertise of the whole Engineering Team;
• Successfully restructured the Engineering organization for improved collaboration, focus, and velocity;
• Successfully led the MVP for building an LLM-chat-based UX to help customers identify their bottlenecks, what they’ve worked on, etc. like their personal assistant. It was an AI agent based on GPT3.5 and GPT4 and used ReAct (reason + act) framework, self-critique, etc.

Led multiple experiments and initiatives to improve both parts of the stack and the processes:

• Built PoCs for improving the data model using event-sourcing and Druid;
• Improved the Python API performance by preloading the DB data in memory;
• Introduced Storybook and Cypress as part of the front-end stack for better component re-usability and testing.

Led the distributed Engineering Team in building a product in the Software Engineering Intelligence space from day 1.

• Leading the distributed Engineering Team (5 persons);
• Coordinated with the Product Manager and CEO to prioritize work in order to be fast and goal-oriented;
• Analyzed feasibility and estimation with the whole Engineering Team;
• Coordinated the Engineering Team’s work to ensure we’re heading in the right direction, ensure and foster good communication, trust, and collaboration;
• Made individual contributions;
• Had the 1:1s with everyone in the Engineering Team;
• Had the last call on Engineering decisions and be the main responsible and accountable.

Achievements:

• Successfully lead the Engineering Team into delivering the first MVP of the product in 5 months.

A system for fetching GitHub metadata that supports multiple providers including GitHub and Bitbucket.

• Was responsible for leading the Applications Team (3 persons);
• Was the bridge between the team and the VP of Engineering, and Product team;
• Coordinated the design of new projects and features eventually along with the other teams be the maintainer of most of the repositories;
• Planned the team’s work in the kanban board aligned with the priority given by the Product team.

Achievements:

• Successfully added important features to metadata-retrieval for fetching Github metadata, such as support for multiple tokens, support for multiple organizations, parallel downloads, etc;
• Successfully built a high-performance metadata retrieval service in Go, supporting multiple providers (GitHub, Bitbucket) through both REST and GraphQL APIs.

An open-source CLI to query git objects in local repositories by gluing together components of the source{d} stack.

• Successfully contributed to maintaining and improving the product called engine (open-source) that mainly provides a CLI to query git objects in local repositories by gluing together all the components of the source{d} stack;
• Successfully contributed since the inception of the new product sourced CE (evolution of engine), composed of the sourced-ce and the sourced-ui projects (both open-source);
• Successfully prepared and delivered demos and proofs-of-concept for both investors and potential customers, showing the potential of the product at its maximum. This required also data visualization skills for the preparation of the dashboards.

Early POCs to experiment different product directions based on data-ingestion in Python and Scale using Apache Spark

• Designed, experimented with, and deployed different proofs-of-concept using different technologies;
• Worked on a core system written in Scala with some parts in Python and based on Spark and its ecosystem.

Responsible for rewriting from scratch the whole analytics pipeline with real-time streaming and exactly-once semantic for critical domain like billing with very high sustained volume

• Maintained and improved the AdServer serving hundreds of millions of ads a month by ensuring scalability, high availability, and low latency;
• Re-implemented the analytics ETL pipeline from scratch, decreasing the time-to-query of an event from more than 20 minutes to dozens of seconds;
• Built a pipeline able to crunch billions of events a day in real time thanks to horizontal scalability;
• Re-implemented the billing system and made it real time by consuming events from the ETL data pipeline with exactly-once data delivery semantics;
• Implemented a mechanism to run experiments by serving different assets on a selectable target of the traffic;
• Analyzed experiments using data visualization and statistical hypothesis testing;
• Implemented basic machine-learning-based components such as sentiment analysis on YouTube comments and a recommendation system for content selection;
• Worked across a Python-based stack using different web frameworks and databases depending on the needs;
• Used Kafka as the messaging system, Zookeeper for coordination, and Druid as the main data store;
• Worked on a front end written in React.

A face recognition system running in a single-page web app for extracting, clustering, labeling, training, and predicting faces from videos.

• Built an MVP of a face recognition system running on a single-page web app;
• Built a web app that allowed users to upload a video from which faces were extracted and clustered together;
• Provided a labeling mechanism to label clusters and train a model;
• Enabled trained models to predict faces in another video;
• Served the web app using Flask;
• Performed image and video manipulation using OpenCV and FFmpeg;
• Used OpenFace, Dlib, and SciPy for face clustering and recognition.