I work where messy technical systems become products: protocol mechanics, market structure, AI workflows, launch operations, and the interfaces that make them usable.
SCROLL TO EXPLOREIST / OPEN TO CONVERSATIONSEST. 2026 - v2.0
AMMs / CLMMs / PERPETUALS / AGGREGATORS / BRIDGES / CLOB RESEARCH / PREDICTION MARKETS / AI AGENTS / RISK SYSTEMS / INDEXING / AMMs / CLMMs / PERPETUALS / AGGREGATORS / BRIDGES / CLOB RESEARCH / PREDICTION MARKETS / AI AGENTS / RISK SYSTEMS / INDEXING /
42k
PEAK DAILY ACTIVE USERS
$150M+
VOLUME PERSONALLY OVERSEEN
150x
MAX PERPS LEVERAGE
18→<3s
PERP ORDER LATENCY
200+
CONTRIBUTORS COORDINATED
( 03 ) - SELECTED WORK
01 / 07
Dexlyn Perps
Led the perps surface from pair-level risk modeling to execution rebuild: funding scripts, OI limits, collateral work, and order placement cut from 18s to under 3.
I work close to the machinery: protocol mechanics, risk, indexing, launch pressure, weird edge cases, and the product decisions hiding underneath.
At Dexlyn, the work moved from early AMM experiments into CLMM liquidity, perps, routing, bridge flows, launch operations, and the less glamorous systems that decide whether users trust a product after the first click.
Embersity is the independent version of the same instinct: take a messy idea, get close enough to the technical shape to make real decisions, then turn it into something with a surface, a release path, and a reason to exist.
Introducing GLM-5.3: Built to Code. Ready for Cyber Defense.
- Top-tier coding and agentic capabilities, achieved through post-training on the 743B base model
- A major leap in cybersecurity, setting a new standard among open models
Tech Blog: z.ai/blog/glm-5.3
My security hot take: Sandboxing efforts immediately falter to the law of diminishing returns. And probably little effort should be put into sandboxing, if at all (other than basic infrastructure/operational necessities).
My take is mostly naive, but I did come up with things pic.x.com/TGHGSJGejg& built integrated policy layers, and have cared about responsible ai since ~2018.
The only real countermeasure to not getting paper-clipped is having some effective layer of AI monitoring AI. (LLM as a judge)... But this causes a lot of annoyance and friction... and theoretically we can not achieve 100% alignment thus must come up with some 9's framework.
btw this is not a solution the hundreds of companies attending Black Hat right now are going to figure out. BC - the solution itself is fairly simple. A mythos-grade model needs to be monitored by a mythos-grade model. And then from there you need even more layers so they don't zombify each other.
It's why I think the labs are pushing heavy on auto mode because at some point they might have to flip it on by default without users being able to turn it off