How it works
Two frontiers are still being drawn, crypto and AI. What holds the attention here is the new primitive each one is producing. In crypto it is the vault, a programmable home for capital. In AI it is the agent, software that acts of its own accord. New primitives are where the interesting product work hides. The path ran through ecosystem and grants in Web3; today the work is institutional crypto products and applied-AI systems.
On crypto, the depth goes to what keeps a vault safe: staking and validator economics, the math of leverage and liquidation, the carry that pays or inverts, tokenomics, and on-chain data queried by hand. On AI, the work is built and shipped: retrieval over real data, fine-tuned open models, and agents that transact on their own rails. Beneath all of it sits one pair, judgment about what to build and proof that it works.
Decisions settled by on-chain data, not decks
When a leverage is called the sweet spot, or a strategy is said to survive a depeg, the claim is read off a live query anyone can open and re-run. The SQL is written by hand, on Dune, against real on-chain state. Opinions are cheap. A reproducible number is not, and it is what gets a risk committee to a yes.
Deep on DeFi mechanics, not just the narrative
Staking and validator economics, the looping math behind leveraged staking, loan-to-value and liquidation thresholds, how carry inverts when borrow demand spikes, tokenomics and governance. The work happens at the level where the product actually lives or dies, then it is packaged so a fund can hold it.
Applied AI that ships, not just prototypes
Retrieval over real databases with retrieval scored separately from generation, and fine-tuning of open models (mostly Gemma 4 on a single A100 with LoRA/QLoRA) when prompting and RAG run out of road. The unglamorous middle, grounding and specialization, is what decides whether an AI product is trustworthy or just a good demo.
Agents that do the work, not just answer
The real shift in AI is from systems that answer to systems that act. Autonomous agents are trained on a business's own data and workflows, then their autonomy is scoped to the blast radius. Every irreversible action, a payment over a rail like x402 for instance, sits behind hard guardrails: approvals, allowlists, spend caps, an audit trail. Crypto is one place they run, not the limit. The discipline is what makes any of it shippable.
For AI, the eval is the product
A model that is confidently wrong is worse than no model. So the eval suite comes before the feature: happy path, edge cases and adversarial inputs, scoring the whole agent trace rather than just the final answer, with passing treated as the price of shipping.
Compliance carried inside, not bolted on
In crypto, the regulatory frame (MiCA in Europe) is a design input from the first sketch. Carry it early and it stops being friction. Nine times out of ten it is the thing that lets a product reach an institution at all.