LOADING
Loading the next screenLOADING
Loading the next screenThe questions I actually get asked about token standards, forward-deployed AI engineering, and running AI agents inside regulated finance. Each answer links to the page or post that carries the full working.
ERC-3643, also called T-REX, is the Ethereum standard for permissioned security tokens. It puts compliance inside the token: every transfer asks an on-chain identity registry whether the receiver holds the claims the token requires, and reverts if it does not. Its status is Final, and more than $32B in assets is tokenized on it. I am one of its five named authors.
The full ERC-3643 explainerThey put compliance in different places, so the answer depends on the shape of your problem. ERC-3643 enforces identity at transfer time inside the token itself. ERC-1400 splits a holder balance into named partitions with their own rules, plus a controller that can force a move, and it assumes something upstream already decided who is allowed on the register. Choose ERC-3643 when the regulator cares who holds the token and eligibility can change. Choose partitions when your instrument genuinely has tranches.
ERC-3643 vs ERC-1400 vs CMTAT, comparedMore than $32B. That figure is the standard's own ecosystem headline rather than a number I measured, and it moves, so read it as a floor and not a live counter.
What ERC-3643 is and where it is usedERC-6960 is a dual-layer token standard. One contract represents both a main asset and fractional shares of that asset: layer one is the asset, layer two is a fraction of it, so you do not need separate contracts or bridging between them. It is a Draft. I am its first author, one of six.
The full ERC-6960 explainerERC-7410 is a small ERC-20 extension. It lets a spender lower or revoke the allowance an owner granted it, without waiting for the owner to submit that change on-chain. It cannot be used to increase an allowance. It is a Draft, and I am listed second among its three authors.
The full ERC-7410 explainerIt goes through the EIP process: you write the spec, it gets a number, and it only reaches Final after surviving public review. The technical writing is maybe 20% of the work. The rest is the community argument and rewriting your spec after someone tears it apart in public.
The anatomy of an ERC, start to finishA forward-deployed AI engineer builds and ships AI systems inside the customer's environment, against their real data and their real constraints, instead of handing over a design and leaving. In regulated finance that changes the job: the audit and compliance requirements are part of the engineering problem from day one rather than a review at the end.
A practitioner definition of the FDE roleRather than answer from opinion, I collected the data: 119 agent and AI-engineer job postings across 32 employers, from public ATS boards. Four in five of those postings name an explicit agent framework or agentic pattern, which is the clearest evidence that this is its own engineering role and not a relabelled AI job. The dataset and the method are both published.
The State of Agent Engineering, with the datasetOnly under containment. The pattern I ship is least privilege plus a propose-then-approve step: the agent can prepare a payment, it cannot execute one on its own, each capability is scoped as narrowly as the task allows, and every action lands in an audit trail. An agent that can move money unattended is a design decision you will have to defend, so I do not make it.
Least-privilege agents that move moneyYou verify the citation instead of just attaching it, and you let the system refuse. Every sentence carries the source span it came from, an entailment check asks whether that span actually supports the claim, any figure has to match the source to the digit, and a single unsupported claim abstains the whole answer. A refusal you can act on beats a fluent answer you cannot trust.
Verified citations and abstention in RAGThis is engineering guidance, not legal advice, and your obligations depend on your system and your counsel. In practice the parts that reach an engineering team are about evidence: logging what the system did, showing that a human can oversee and override it, and being able to demonstrate accuracy rather than assert it. Those have to be built in. Retrofitting them is the expensive path.
The EU AI Act as an engineering checklistWhen an eval suite that is allowed to say no says yes. Score the dimensions that actually fail in production, put a release gate on them, and feed red-team prompts into the same suite instead of running them as a separate exercise. If nothing in your process can block a release, you do not have a gate, you have a dashboard.
The eval suite that gates a releaseI am an AI and agent engineer for regulated finance, based in France, and a former CTO. I have authored three Ethereum standards and shipped production multi-agent, RAG and LLM systems inside regulated organizations. The long version, with dates and receipts, is on the about page.
About, with the full recordThose are different questions and they have their own answers: what an ERC-3643 audit covers, what I currently have capacity for, how engagements are structured, and what happens if it does not work out.
▸ Read the services FAQAsk it directly. If it is a good one, it ends up on this page.
▶ ASK MEPrefer email? contact@adam-boudjemaa.com