Staff Engineer · data platform and applied ML
Hi, I'm Roberto.
I'm a Staff Engineer at Audiense. I've been building software since 2011, and I enjoy it most when the problem is new to me and I have to learn my way into it. I like to keep the product in view: what a team or a user actually needs matters more to me than the technology I use to get there. I'm curious by nature, and most of what I know I picked up by building things myself.
At Audiense I lead the data team technically. I'm responsible for how the company's data is stored, queried and kept fresh: the lake; VEGA, the query API product teams use instead of keeping their own copies; and the models that tag profiles as data arrives. I also run internal sessions on agents and fine-tuning, and built an agent skill other teams use to write VEGA queries. The details are on the projects page.
On my own time I build things to learn how they work, and one of them, Radar, is now a live pilot. I'm from Cuenca and I have two small kids, so most of my time outside work goes to them. When it doesn't, it goes to the gym or to the countryside.
Selected work
At Audiense
Replacing a 32B GPU model with a fine-tuned small one on CPU
A fine-tuned GLiNER on CPU replaced Qwen2.5-32B for entity detection at an order of magnitude lower cost. Competitive, not equivalent: it trails mainly on brands, and I shipped it knowing that.
GLiNER · EKS · Redpanda Applied MLInterest classification: keep a category only if two models agree
Assigns interests from a 444-category taxonomy, with a LoRA-tuned bi-encoder and a cross-encoder that must both rank a category in their top 5. Human-judged strict precision went from 36% to 52%.
LoRA · Cross-encoders · vLLM PlatformOne query API instead of a copy per team
VEGA is an asynchronous query API over the lake. Product teams retired at least six local copies of the data, along with several ETLs and a costly Spark process.
FastAPI · Redis · RedpandaOutside work
Personal projects
- Radar — public tenders for small firms, filtered and sent as a weekly email; launched in September 2026 as a pilot (licitacionespyme.es).
- Hermes Expense Tracker — shared household expenses over Telegram; a FastMCP server, SQLite, one assistant per person.
- Voice agent prototype and math tutor — STT → LLM → TTS on LiveKit, with explicit state machines in one and a validation harness in the other.
- AI research wiki — an Obsidian vault on agents, tool use, RAG, evals, observability and voice, maintained with Hermes Agent.