Software Architect at Synergetics.ai. Backend and platform work, with a specialty in custody: who is allowed to move value, and how you prove it. For a few years that meant humans holding keys. Now it means agents holding their own wallets, which is a harder version of the same problem.
Before Synergetics, wallet and key-custody engineering at NetObjex / Kresus Labs.
Multi-tenant platform foundation — Synergetics
Postgres row-level security with composite tenant foreign keys, one sanctioned bypass path
rather than none, and a transactional outbox so events and state commit together. Where
isolation lives is the whole decision. Put it in application code and one missed WHERE
clause defeats it; put it in the database and you have to design the escape hatch on purpose.
Regulation compiled into the system, not written into a policy — Synergetics Outbound messaging that has to satisfy CAN-SPAM, CASL, GDPR, DPDP and PECR at once. Rather than a compliance document nobody reads, jurisdiction is a routing decision in code: consent model, retention, and unsubscribe handling resolve per recipient before anything sends. Legal constraints are easier to keep when the system can't violate them.
Agent systems in production — Synergetics Multi-agent workflows on Claude, MCP tool servers, and the parts that make them safe to run: retrieval that has to cite its source, a fact-guard that blocks any figure not present in the source of record, and approval gates enforced server-side rather than in the UI. Most of the design work in agent systems is deciding what the model isn't allowed to do alone.
Agent identity and wallets — Synergetics How an agent proves who it is and what it may spend: ERC-8004 registries, W3C DIDs and verifiable credentials, ERC-4337 smart accounts, ERC-6551 token-bound accounts so the identity NFT owns the wallet. A2A covers agent-to-agent task delegation and MCP covers tools; the trust layer sits under both, and that's the part nobody has settled.
Account abstraction and key custody — Kresus Labs An ERC-4337 wallet: smart accounts, account factory, and the paymaster service behind gasless and sponsored transactions. Then the threshold-signing relay for the vault, coordinating multi-party signing rounds so no single node held a whole private key. Worth being precise about what that buys you. It removes the point where a complete key exists. It does not remove custodial risk: hold t of n shares and you have custody, and the attack surface moves to whatever authorises the signing.
APIs and SDKs — Kresus Labs 20+ backend services across wallet, vault and marketplace, and three client SDKs so internal and partner teams could integrate without touching chain-level code. The SDK work taught me more about API design than the APIs did — you find out what your interface really costs when someone else has to live inside it.
Things I don't have clean answers to yet:
- An ERC-6551 token-bound account isn't an ERC-4337 account. The 6551 spec never mentions
4337. So if an agent's identity NFT owns its wallet, something still has to implement
validateUserOpbefore a bundler will touch it, and I'm not convinced that belongs in the TBA itself. - Session keys are the piece agent wallets actually need: scoped, time-boxed, spend-limited authority instead of handing an agent an unbounded key. Working through ERC-7715, ERC-7710 and the 7579 module story now.
- ERC-8004's reputation registry is the part I trust least. Anything permissionless and reputational attracts Sybils, and the early on-chain data suggests it already has.
- EIP-7702 shifted the account-abstraction picture more than most people have absorbed. Still forming a view on where it leaves 4337 for existing EOA users.
I keep a decision log. Context, the options I had, what I picked, what it cost, and what would make me change my mind. The one for my current platform work is 128 entries. When a decision gets superseded I add a new entry pointing back rather than editing the old one, so the log stays honest about what I believed at the time.
I lean on Claude heavily for the mechanical half of the work and review everything that ships. Building agent systems for a living has mostly taught me where not to trust one.
TypeScript, Node, Fastify, Express · Postgres, row-level security, transactional outbox · Docker, GitHub Actions · Solidity, Hardhat, Slither, solhint · Ethereum, Base, Polygon PoS · Anthropic Claude API, MCP, A2A