Introducing Edgard ..... a PM who ships APIs banks can audit
Why Now
Wed Sep 16 08:00:00 2026Day 1 since Jev went public
Counting from the announcement of System One models and Jev on September 15, 2026.
Five roles open on the TypeSafe board. Product ownership of the API is described inside two of them and posted in none.
"partner closely with product managers and customers"MTS, Model Capabilities
"abstractions that will serve us ... as our billing needs and business logic needs grow"MTS, Backend/Platform
Out Of Stealth
TypeSafe announced System One models and Jev on September 15, 2026, after two years of quiet work. There is a waitlist, a published price of $42 per billion input tokens, and a claim: typed decisions with calibrated confidence. What there isn't yet is a customer who has put Jev inside a production workflow and let it act on its own. The first ten of those set the pattern for everyone after, and they decide what the API has to look like.
Five Roles, No PM
The open roles are three engineers, a marketer and a blank "Member of Staff". The Model Capabilities post already expects engineers to partner with product managers and customers. The backend post asks for API abstractions that hold up as billing needs grow. Both assume a person who owns the customer side of the API, from design partners to pricing. That seat is empty, and I would like to be the one who fills it and defines it.
The Buyer Is A Risk Team
The obvious first customers for a cheap, fast decision model are back offices in finance: reconciliation breaks, transaction categorisation, document classification, exception routing. Those buyers do not open with accuracy. They open with three questions. What happens when confidence is low, who reviews it, and can I see every decision afterwards. I have sat across that table for years, at Croesus and at Desjardins, and I know what they need to see before anything ships.
How I'd Help In The First Months
62 / 80 decisions acted on their ownillustrative run, one decision type
The thresholds are the product. A customer sets them per decision type, and the log proves they were respected.
Weeks 1 to 4
Design Partners
Pick three to five workflows where a wrong decision is cheap to catch and a slow one is expensive: reconciliation breaks, transaction categorisation, document classification, exception routing. Sign design partners from the waitlist, wire Jev into one step of their pipeline, and measure two numbers against the chat model they use today: cost per decision and escalation rate.
At Croesus every new custodian brought its own settlement rules and its own users, who would notice within a day if a number was wrong. That is the standard these partners will hold Jev to, so it is the standard the pilots are built around.
Weeks 2 to 8
The Trust Layer
Own the parts of the API a risk team reads before an engineer reads the docs: a decision schema per use case, threshold settings that say when Jev acts and when it hands off, a review queue for the hand-offs, and a log of every decision with its confidence and its outcome.
My trading bots run unattended inside thresholds I set, with automatic halts when a limit is crossed. Same shape, different asset. I have also shipped access controls, identity flows and audit trails for banks, which is the vocabulary their model-risk teams will use on the first call.
Weeks 6 to 12
Pricing & Packaging
Work with the backend team on the billing abstractions their own job post already asks for. Metering per decision rather than per token is what makes the $42-per-billion story legible to a CFO, and a threshold-aware price (acted, reviewed, declined) is a story nobody on chat pricing can tell.
Ship one self-serve tier and one design-partner agreement, then let usage decide the third. I built the pricing and packaging conversation at SweetIQ from product data and customer interviews, and I would run it the same way here.
Schematic, not data. The point is that the two vertical lines belong to the customer, and moving them is the whole product conversation.
What You'd See At Day 90
Who
$1.5M
New annual recurring revenue from a multi-custodian platform: integrations, provisioning, settlement and reconciliation, access controls, identity flows, audit trails.
70%
Less time to generate criteria, from an NLP search I chose the model for, set the latency budget on, and designed so reviewers would trust it instead of route around it.
38%
Faster delivery across 20+ Product Owners and designers once the practice had standards, rituals and story-quality frameworks. NPS up 35 points in the same period.
50%+
Lower operating cost after replacing manual back-office workflows with an API-driven platform. Support volume down 35%, SLAs 200% faster.
Ten years of product work, most of it inside banks and the firms that serve them: Croesus, Desjardins through Alithya, LCL Private Bank. The pattern repeats. One platform has to serve many institutions, each with its own rules, its own permissions and its own auditors, and the job is to absorb that as configuration instead of building one integration per customer.
Since 2025 I run EMB: systematic strategies across derivatives, futures and on-chain perpetuals, and arbitrage bots that work unattended across EVM chains, in Python I wrote. The thresholds that stop a strategy are the part I think about most, which is why Jev's confidence model reads to me like a product I have already been using from the other side.
Canadian and French citizen, based in Montreal. I would relocate to San Francisco for the in-person office, and I would need the visa sponsorship the posting lists. I have read the manifesto and the Bitterest Lesson post, and I agree with the premise: compute is cheap, doing the right task is the hard part.