Juan Rosero
I'm a Product Engineer working in web platforms since 2018. I spent most of those years at Engine AI where I grew from Junior Developer to Squad Lead. I started my career in Ecuador and made it all the way to London thanks to my interest in solving problems, my communication skills and a keen eye for a good design. Most recently focused on developing AI products that bridge the gap between what the models can do and what the people actually need them for. Right now I'm leading the squad behind one of the most important projects for Engine AI, a RAG-based research assistant that serves 1000+ users across a collection of about 20,000 documents. I've decided it's time for my next challenge, so I'm handing it over properly before I go.
// STACKI joined Engine AI at their Lisbon offices as a Junior Full Stack developer for the dashboard engine: a platform that allowed users to build data-rich applications and connect that data to easy to use visualizations. When the company pivoted to the AI space we leveraged that platform and found new paths to pull insights from the client's structured or unstructured data.
The project started as a POC based on one of the templates I helped to build. The goal was clear: our client wanted to navigate their vast collection of documents with ease. That unlocked another cool feature: by analyzing the questions we got asked, we started understanding the consumer side, which could improve the research itself. In this journey my responsibilities covered:
- Build and maintain the architecture that powers the solution. Most recently we moved away from a step heavy process to a more tool based approach using Pydantic AI. The agent went from feeling like a robot that can only answer questions to an interactive assistant that helps you find an answer.
- Communicate with the client on a weekly basis. Simplifying the technical details so they understand the why behind the roadmaps that we had planned.
- Develop the evaluation framework. I had to ask "How does a good answer looks like?" and put the answer in numbers. This was initially built as an in-house solution but moved to Pydantic Evals after the switch to the tool based architecture.
- Actively improve performance, accuracy and transparency. That's where the evals came in clutch allowing us to figure out steps that could be parallelized or delegated to faster model. It also flagged cases where regressions were introduced. As for the transparency bit, we mimicked the Open Responses protocol, so the user knows what the agent is doing in real time and provide easy to navigate links to the sources of an answer.
- Building the infrastructure. One of the requirements was that the data resides on AWS services in the US. I was in charge of picking the right AWS services to deploy the app and to help automate that CI pipeline.
Before working in Research Intelligence, Engine AI was hunting for possible clients. When we wanted to showcase an idea of what was possible using our platform, I was the one to take over those high-stakes demos.
- I built 5 working prototypes, 2 of which helped us close deals with clients, one of them of course was Research Intelligence.
- The other successful prototype was called Product Testing. This was a React app that connected to our widget generating agents and rendered them in the client's branding. This started as an internal testing tool, hence the name, but evolved to form part of client facing presentations.
When I joined the company, the Dashboard Engine had great foundations but needed to be extended to reach it's full potential. The job was centered around our widgets and the system that bonded them together.
- I created the GraphQL API and definitions that allowed users to create widgets, layout them and connect them to data sources.
- Developed connectors using Duck DB that allowed the backend to query them and eventually pass the data to the widgets in the Frontend.
- Built the React Component Library that powered different visualizations, mainly using Highcharts.
- Worked with the design team to develop our proprietary design system. This allowed us to show our product in the client's own branding.
Technically proficient. I can jump into a complex codebase or problem with limited context and figure it out. I ask the right questions, make informed decisions where I need to, and try not to let uncertainty become a blocker for the people around me.
Good communicator. I meet people where they are. I listen more than I talk, and I want people to feel safe being honest with me. That said, I know when the moment calls for clarity and I don't shy away from saying the important thing.
Design mindset. I love the craft of making something that just feels right to use. I work closely with design teams and take the experience people have seriously, and continuous iteration is how I get there. My favourite kind of design is the kind people don't notice.
- Full merit scholarship awarded by the State
- Thesis: Bottle image classification system using Convolutional Neural Networks