Product leader — AI systems, crypto, DeFi. Barcelona.
Head of Product, DeFi at Kiln through April 2026, where I launched OmniVault and Railnet with CoinShares and Monarq as institutional launch partners. Before that, Director of Product at 0x / matcha.xyz — open-source protocol to a 600-customer API business. Before that, Ads at Facebook. Aerospace engineer by training.
A multi-model LLM agent fleet on my own hardware, dispatched over Telegram by voice and text. Since February 2026:
- ~3,500 briefs dispatched · 8,000+ documents produced · 5.1B tokens
- 12 models, tiered. 63% of runs mix tiers in a single pass; the cheap tier carries that 63% for 3% of total spend
- $0.79 median brief. Instrumented to the unit cost of a single answer — which is how I found that cache reads are 94.3% of input tokens at 93.8× amplification, and that the expensive part of an agent is not retrieval but retention
It writes its own memory — an append-only journal compiled into ~220 entity pages, loaded at every dispatch — and runs a nightly reflection over the day's work, then weekly and monthly consolidations.
- expatrimony.com — untangles state pensions across borders. Production LLM extraction over arbitrary PDFs: structured outputs, dual provider (metered API + self-hosted fallback), and an eval harness that grades extracted numbers against a tolerance on a weekly cron, deliberately kept out of PR CI because a probabilistic extractor has no business gating a merge.
- loopclub — a 16×4 drum machine that lives entirely onchain, on MegaETH. Ships with an MCP server, so you can describe a beat to Claude and get back a playable link.
- shillscore — signal-scoring for a noisy timeline.