Saito Hiroyuki

Founder @fabrknt | I just want to build products that I want to use by myself.

[Saito Hiroyuki @psyto]

Most Ethereum onboarding treats Ethereum as a separate magical world.

Question: what if the better frame is database + distributed system + compiler + network + concurrency runtime, glued together by consensus?

Principle: the bug classes are not Ethereum-specific. Race conditions, deadlocks, backpressure, VM dispatch, peer scoring. Same systems problems, different battlefield.

https://t.co/eFLJr2k4G3

[Saito Hiroyuki @psyto]

Hyperliquid のエンジンは、外からは誰も読めない。

そこで自分でオープン版を作ってみた — Rust、Reth で。CLOB → ファンディング → 清算 → ADL。

なかなかいい出来だと思うので、それを「自分でソースから作る」コースにしてみました。

https://t.co/ZfoZRAJQAG

[Saito Hiroyuki @psyto]

Nobody outside can read Hyperliquid's engine.

So I built the open one — in Rust, on Reth. CLOB → funding → liquidation → ADL.

Then I turned it into a course where you build it yourself, from the source up.

https://t.co/ZfoZRAJQAG

[Saito Hiroyuki @psyto]

Honestly, Fabrknt feels like it is made for people who already suspect the important part is happening somewhere below the interface.

People who know the polished surface is not the system.

People who want to see where stress goes.
What assumptions hold.
What breaks.
What depends on what.
What kind of machine is actually there once the language gets stripped away.

If that instinct already lives in you, this might make sense:
https://t.co/f19GmWEQbK

[Saito Hiroyuki @psyto]

If you keep books for an on-chain wallet, you probably know the manual version of the job.

Without Intentio, you have to reconstruct the transaction yourself:
open the transaction,
read the logs,
trace the transfers,
figure out what actually happened,
then turn that into one ledger line by hand.

With Intentio, that reconstruction is automated.

You paste a public wallet address.
Intentio classifies the wallet activity, reconstructs the economic event, and gives you a ledger you can actually review.

It stays read-only.
No wallet connect.
No signing.
And when the shape is unclear, it abstains instead of mis-booking.

A simple swap can emit 11 Transfer logs even when the thing you actually need to book is just 1 economic event.

That is the work Intentio removes.

https://t.co/tdx51SXO2a

[Saito Hiroyuki @psyto]

If I had to compress the Hyperliquid lesson into one sentence:

people thought they were competing on app quality when they were actually competing on market structure.

That doesn’t mean UI doesn’t matter.
It does.

It means UI was never the whole game.

The hard part is that market structure is less legible.
It’s harder to explain.
Harder to screenshot.
Harder to market.

But when it’s right, it compounds.
And when it’s wrong, no amount of front-end polish saves you.

I think that’s one of the most important lessons in DeFi right now.

[Saito Hiroyuki @psyto]

My current take:

Hyperliquid won because traders could feel the market structure.

Not in an abstract whitepaper sense.
In the literal sense of:
execution quality
book behavior
latency feel
venue coherence
microstructure trust

That matters more than crypto people like to admit.

A lot of teams are still building “perp products.”
Hyperliquid built a trading machine.

That difference is everything.

[Saito Hiroyuki @psyto]

For the curious:

Trigger a Treasury shock. Watch four
Megacorp prime brokers liquidate in
real time.

Cinematic UI. Tokyo 2042.

→ https://t.co/74jpGUB4SF

[Saito Hiroyuki @psyto]

Open Rust DeFi L1 stack + source-grounded
curriculum + Tokyo 2042 cascade simulator.

Three venues, one engine. Pick yours ↓

[Saito Hiroyuki @psyto]

For evaluators:

Reth-based perp DEX engine, 10 scenarios,
each emitting HEADLINE / TIMELINE / DELTA /
OUTCOMES / NEXT — runnable from fresh clone,
one cargo run.

Solana perp + prime broker engines next.

→ https://t.co/q9Ro0OQDEv

[Saito Hiroyuki @psyto]

If you're building a margin engine, a matching engine, or any L1 state machine: start with the properties, not the types. The types follow.

What I'd tell day-1-me: don't write the engine until you can state the invariant it preserves.

Code: https://t.co/riV19mwppK

[Saito Hiroyuki @psyto]

When I wrote the OpenHL liquidation engine — 4 states, pure-compute, no I/O — the very first proptest I ran failed.

Not failed-as-in-edge-case.

Failed-as-in "my margin math doesn't survive leverage."

[Saito Hiroyuki @psyto]

The thing nobody warned me about pure state machines: the discipline of writing them is the discipline of writing properties strong enough to refute them.

The state machine is small.

The property test suite is bigger than the engine it verifies.

Worth it. Not free.

[Saito Hiroyuki @psyto]

In OpenHL this gave me two precompiles:

CLOB_READ_BEST_BID = 0x...0c1b (read book from Solidity)
CLOB_PLACE_ORDER= 0x...0c1c (place orders → matching engine)

Solidity talks to a Rust matching engine at execution speed.

Full walkthrough on RethLab: https://t.co/gI4IOvMUDa

[Saito Hiroyuki @psyto]

Reading openhl_evm.rs again this morning.

Always struck by how compact Reth's EVM customization surface is.

To build a Hyperliquid-shape L1 — perp DEX with custom matching engine, precompile-driven design — the whole hook is one trait + one swap point. ~80 LOC.

[Saito Hiroyuki @psyto]

What I find load-bearing about this:

register the precompiles once and they show up in every EVM Reth instantiates — payload build, block validation, `eth_call`, `debug_trace`, anvil fork harnesses.

Zero per-callsite wiring.
That's the abstraction HL-shape systems need.