About

I'd been shipping ML in production since the Oony days, and watching the language-model space closely from early on — but it wasn't until late 2023 that the agentic side started feeling like 1995-vintage internet again: primitive, but obviously load-bearing. I went all in. Production work with LangChain and LangGraph, a lot of time on retrieval and evaluation, and a growing conviction that agents and Web3 wallets are the same problem viewed from two angles.

That has been the through-line since: first inside Hiero, and more recently working independently with startups and enterprises to move AI initiatives from advisory into production — LLM applications, advanced RAG pipelines, domain-specific fine-tuning, and agentic workflows. The recurring shape of it is expanding automation from single-step LLM calls into stateful, multi-step execution: structured outputs, tool calling, retrieval, long-term memory, and the orchestration layer holding those together. Most engagements start somewhere in the gap between a demo that convinces a room and a system that can be trusted to run unattended.

What I'm building now sits on the same seam. zkStash is decentralized memory for AI agents. Different codebases, same runtime problem.

I joined Hiero as founding engineer and CTO, leading product and the early team, and wrote most of the initial codebase myself. The premise was a complete AI agent terminal with real blockchain access across Solana and Base — and we started building it before "agentic" had become the term of art.

The surface was wide: agents that could actually act on-chain — access DeFi primitives, create and launch tokens, buy, trade, and analyze tokens and NFTs, run deep price research, and chain those actions into recurring tasks the agent executed on its own schedule. The fun, hard part was the runtime — multi-step research-and-execution graphs on LangChain, LangGraph, and LangSmith that planned across chains, called the right tool, and stayed observable. Later the product grew past conversational assistance into real-time blockchain event monitoring and asynchronous execution, which meant event-driven services running underneath the agent graphs rather than alongside them. We shipped the platform and the HTERM token. The broader agent-token cycle cooled and the project eventually wound down, but it was the deepest stress test I've run on what production agents can actually do once you hand them a wallet, a toolbox, and an audience.

I founded TradeStars to test a thesis: sports fandom is already a market — fans value players, argue about price discovery, and trade reputations on group chats every weekend. What if that market were real?

We built it. Each athlete became a fractional NFT — an ERC-20-backed share of an NFT, priced by an automated market maker that responded to real-world performance. We launched the public exchange in January 2020, ran it on Polygon for the L2 economics, raised $2.2M+ across rounds, and shipped the TSX token. The fractional-NFT primitive we designed showed up in a lot of places afterwards.

My route into crypto started with provably fair gaming and state channels: games where the important question was whether two parties could move fast off-chain while keeping settlement honest on-chain. It felt like a natural continuation of the systems work I had always liked — protocol design, adversarial edges, and software that could enforce rules without asking a platform to be trusted.

I joined Decentraland in its early days as tech and scalability lead, inside the initial fifteen-person team, while the protocol and the marketplace were still being shaped. I co-developed one of Ethereum's earliest ERC-721 marketplace contracts and the Web3 transaction flow around it; it went on to settle roughly $4M in transactions. The other half of the job was scalability research — prototyping Layer 2 approaches for cheaper, higher-throughput transactions, which brought me back to state channels and, for a while, plasma chains.

It's where I first saw NFTs as a primitive worth building real applications on, not just a collectible format. I left to start TradeStars, but the conviction I picked up there — that on-chain ownership of digital things would matter — set the next decade of work.

With my long-time co-founder Ariel Barmat I built Weegoh, one of the first location-based social networks in Latin America, and then Oony, a daily deals and shopping recommendation platform that grew to a catalog of several million products. Over six years we took it to 1M users across fifteen countries and 4M monthly visits, generating $2M in annual revenue — which meant building the localization architecture and the data infrastructure to survive that many launches. Oony is also where I first shipped machine learning in production — classifiers for deal quality and ranking, back when "ML" still meant scikit-learn and feature engineering. The thread that started there runs straight through to what I'm building now.

I came up writing C, working close to the metal — networks, routers, packet handling. In 2005 I created the SIP connection-tracking and NAT modules for the Linux kernel's netfilter subsystem — nf_conntrack_sip.c and nf_nat_sip.c. They handle the awkward parts of routing SIP signaling and media through NAT, where addressing information is buried inside the protocol payload itself. That code shipped into the mainline kernel and remains the canonical SIP conntrack/NAT implementation in Linux. SIP itself remains the signaling backbone of much of the world's voice and video calling, from VoIP services to modern carrier networks. There's a short design write-up on the netfilter site from when it landed.

Conversations with founders building agentic products, infra teams working on agent identity / payments / memory, and operators who need a technical co-founder or architect-level partner for the next stretch. If that's you, say hi.