Product Lead · AI systems, zero to one
Two AI systems shipped to production this year - built spec-driven with PRD, glossary and ADRs, coding agents doing the typing. Founding product manager at Proxity; before that, scaled a legal-media platform to 150K MAU and $500K ARR.
An internal due-diligence engine for the group's investment team: 13 stages, cited facts, human review. Agentic loops only in 2 of 13 stages - per-stage LLM cost ~3x lower. Alongside it, I run the Proxity platform itself with an external dev team.
A self-enriching market-intelligence graph for asset deals: an AI pipeline drafts, a human editor decides, and every decision feeds back into the graph. Public launch planned for Q4 2026.
A catalog that fills itself from a Valencia agency's Telegram channel: 13 filter groups, a map, one-tap contact with the agent in 4 languages, an admin panel. Deterministic parser, LLM only at the edges.
Built the product organization from scratch, monetized the B2B side with a custom CRM, then stayed on as a part-time consultant shipping AI automation across the operation.
Designed the methodology, personally ran the interviews, and turned raw market data into a top-30 ranking and a public B2B analytics platform for international investors.
Analyzed how lawyers actually worked, then shipped document automation, analytics API integrations, and a knowledge portal for 200+ employees.
I'm a Product Lead who takes products from zero to production and into the market - lately, AI systems.
Right now I'm the founding product manager at Proxity, an invite-only deal platform for confidential sales of large commercial assets. I own the platform end-to-end with an external dev team, and I designed and built its internal AI due-diligence engine: a 13-stage pipeline on FastAPI, Pydantic AI and Google Cloud that turns 2-3 analyst-days into a source-cited investment-committee report in under 30 minutes, with human review. For IInsider, a sister project in the group, I took the founder's concept to an AI pipeline and knowledge graph of the deal market - from PRD to unattended production in 11 days.
I build spec-driven with parallel AI coding agents: a versioned PRD, a domain glossary and architecture decision records are the agents' source of truth; scope, architecture and review stay mine.
Before that, I scaled a legal media platform from concept to 150K MAU and $500K ARR as Head of Product & Operations, and I still advise it part-time as a Product & AI Consultant. On the side, I built and sold a searchable listing catalog to a Valencia real-estate agency, solo, in 3 weeks from prototype. My foundation is 5+ years in Russian insolvency and banking law, which taught me to look for the structure in a messy process and to model complex domains end-to-end.
What I care about: AI systems that do real work under human review, domains with complex rules, and the boring infrastructure that makes powerful tools actually usable.
Based in Valencia, Spain, with Spain work authorization.
The spec-driven system behind 70-80% agent-written code with 100% human-owned architecture and review: PRD, glossary and ADRs as the agents' source of truth, and honest numbers on what changes.
One of four talks I gave at QA Breakfast Valencia in 2025-2026, three of them on AI. This one closed the June 2026 season at RingCentral: what AI-speed building does to product judgment - argued through a feature I shipped in an evening with zero review changes, and a feature engineers built where I found 9 product and UX issues.
Three merged pull requests to the AI reading-assistant plugin for KOReader: dictionary with sentence context and translation, an AI dictionary gesture on text selection, and a model picker for OpenRouter. Started as reading sci-fi in English, ended as shipped features.
Type a job title and see which skills employers list most often, across 1.15M unique LinkedIn postings from January 2024 - built in 2025 after LinkedIn stopped showing this itself. Rebuilt and open-sourced in 2026 as Streamlit and DuckDB over Parquet on Cloud Run, with whole-word search that handles C++ and .NET. Live demo. Next: fresh 2026 postings, with skills extracted from job descriptions.
A Firebase app that turns a rough research question into a precise Deep Research prompt through AI-generated clarifying questions - a pre-research step that saves 10-15 minutes of prompt tweaking per report.
Hands-on notes on prototyping and applied AI: prototyping with Firebase, what got cheap and what got expensive in AI-speed building.