Baptiste Castiglione
I'm a product manager working across two frontiers, crypto and AI. I turn the new primitives each one is producing into products serious people can actually use.
In crypto, I focus on vaults: the emerging primitive for how capital is held, governed, and put to work on-chain. In AI, I focus on agents: software that has moved from answering questions to taking actions, and now pays its own way. Two domains, one job.
The discipline carries across both. In crypto, nothing ships unless on-chain data backs it. In AI, no agent runs unattended until it can be trusted with money. Below is what I build and how I think about it.
What I do
Crypto vaults
The new primitive for capital on-chain: non-custodial, rules baked in, risk bounded. Staking loops, leverage, liquidation, carry. I write the SQL myself, so every number has a query behind it.
AI agents
Software that acts instead of just answering, and now pays per request over rails like x402. The product work is control: spend caps, allowlists, approvals, and evals on the whole trace.
Product, zero to one
I find the job nobody has done yet, then build the shortest path to it. Discovery, positioning, roadmap, and the unpopular call on what to cut.
Proof and compliance
Decisions settled by on-chain data, not decks. Regulation treated as a design input, not a tax paid at the end. That is often what lets a crypto product reach a serious institution.
Selected work
Leveraged Staking Carry Index
A live dashboard that answers one brutal question: at what leverage does looping staking still pay, and where does it blow up. Read off on-chain data, not vibes.
Leveraged Staking Vault
Turns the thin staking spread into something a fund can actually hold. The client gets the yield. They never touch a DeFi protocol.
Validator Distribution Dashboard
See and steer how stake spreads across validators. Performance, concentration and health, in one glance.
Autonomous Agents and Machine Payments
Agents that act, fetch, and pay their own bills over x402. The product work is not the intelligence. It is the financial brakes that keep an autonomous spender from being insane.
Applied AI: RAG and Fine-tuning
I fine-tune open models, mostly Gemma 4, on a single A100, and build RAG over real databases. The unglamorous layer that decides whether an AI product is trustworthy or just a good demo.
Toolbelt
Take a break
Space, up arrow, or tap to jump. Yes, I built a Chrome-dino. A product is not finished until it is a little bit fun.