Founding AI Engineer | Python · TypeScript · Node · Temporal · RAG | Agentic Transactions at Scale |

Job Description

Yes move from NYC to SF ~~

\n

\n

"Never tell me the odds."

\n

(Han Solo, The Empire Strikes Back)

\n

\n

The odds

\n

\n

A category of global commerce worth trillions still runs on email chains, scanned PDFs, spreadsheet macros and one person who remembers which supplier actually has stock. Every serious software company has tried to improve this from the outside by selling tools to the incumbents. The tools get bought. The work stays manual.

\n

A San Francisco startup has decided the only way through is to stop selling software to the operator and become the operator. They have just closed an inception round in the tens of millions from one of the most recognisable venture firms on earth to do it.

\n

So: long odds, enormous prize, and a founding team that has already flown this corridor once.

\n

\n

Who is behind it

\n

\n

The founders built a previous company serving several hundred enterprise customers who move billions of dollars of inventory every month. They spent years watching precisely where value leaked out of that process, documenting it, and being unable to fix it from where they sat. The engineering and product leadership have come across with them. They know the customers by name, they know the data, and they know every failure mode they are about to engineer around.

\n

That matters more than it sounds. Most early-stage teams spend eighteen months discovering what the problem actually is. This one started on day one with the map already drawn.

\n

\n

What they are building

\n

\n

An AI-native platform that executes transactions rather than assisting with them. Agents interpret a messy request, identify the correct item, locate and price supply, place and confirm the order, re-source when a vendor goes silent, reconcile documents against what was agreed, and manage settlement. The system does the job. A human supervises exceptions.

\n

Because the platform sits in the middle of live transactions, every deal generates proprietary data that nobody else holds: who truly has stock, what it truly costs, which substitution a buyer will accept. That data compounds into a network effect across the supplier and customer graph. The ambition is to own the category outright rather than take a slice of it.

\n

\n

Five systems someone needs to own

\n

\n

You take a domain and run it from architecture to production to the customer phone call. There is no product manager writing the ticket first.

\n

\n

Entity resolution and canonicalisation. Thousands of catalogues describing identical items in mutually incompatible ways, with duplicate records and inconsistent identifiers. Fuzzy matching, deduplication, and the graph that turns the mess into something queryable. Everything else sits on top of this.

\n

\n

Retrieval-augmented substitution. When the requested item is unavailable, find the right equivalent. Embeddings, vector search, retrieval quality tuning, and LLM reasoning over structured specifications where an incorrect answer ships the wrong component to a live site.

\n

\n

Agentic search and price discovery. Agents that scan a fragmented vendor base, compare options and act. Real tool use, real guardrails, and evals that tell you when quality drifts. Production, not a notebook.

\n

\n

A durable-execution platform. The backbone. Thousands of long-running workflows must survive crashes, retries, timeouts and partial completion without double-placing an order. Temporal or similar workflow engines, event-driven architecture, idempotency treated as doctrine, state machines modelling the transaction lifecycle properly.

\n

\n

High-volume data engineering. Document extraction, transformation and manipulation across unstructured records, feeding everything above.

\n

Stack: Python, TypeScript and Node.js, with Postgres, distributed infrastructure and applied LLM work throughout. Deep production experience in at least one language is essential; strength across all three is ideal. No domain knowledge required, and some of the strongest people there arrived with none.

\n

\n

"The ability to speak does not make you intelligent."

\n

(Qui-Gon Jinn, The Phantom Menace)

\n

\n

How they read a CV

\n

\n

Worth being direct, because this team screens hard and declines quickly.

\n

They are looking for the person who designed the system, wrote it, shipped it and still carries the pager. Several impressive candidates have been passed over because the work turned out to be product management, programme management, data science or go-to-market wearing an engineering title. Orchestrating agents and tooling without having built the machinery underneath reads as thin here.

\n

\n

They also read for slope rather than seniority. Two shapes convert. The first is early career, two to five years, with exceptional computer science fundamentals and progression fast enough to be visible on the page. The second is the systems architect, up to fifteen years, who has built genuinely complex distributed systems from scratch and now wants founder-level ownership instead of another promotion cycle. Steady competence in the middle has not been getting through.

\n

\n

And they read closely. Company-level revenue claimed as personal impact, a profile headline that contradicts the CV, or consulting engagements framed as operating roles will surface in the project deep dive. Specific, attributable, defensible work is what travels.

\n

\n

Where strong candidates tend to come from

\n

\n

Frontier AI labs and data infrastructure companies set the bar: OpenAI, Anthropic, Databricks, Snowflake, Confluent, and the newer AI-native product companies with serious engineering reputations. Google, Stripe, Palantir, Scale AI, Ramp and Anduril all produce the right shape. Titles that resonate: Founding Engineer, Founding AI Engineer, Member of Technical Staff, Forward Deployed Engineer.

\n

\n

Education is not mandated, though "exceptional school" recurs in how the leadership describes their best hires. Read that as Stanford, Berkeley, MIT, CMU, Waterloo and equivalents, ideally with something remarkable attached: a competition result, published work, an open-source project that found an audience, or a company started while still studying.

\n

\n

Three things weighted as heavily as the code

\n

\n

Autonomy, stated plainly by leadership: run at hard problems without being pointed at them. Customer-facing comfort, which is a dealbreaker rather than a preference, because every engineer speaks with buyers and suppliers directly. And one remarkable thing on your record that makes people lean forward.

\n

\n

The flight check

\n

\n

Four stages, run fast. An opening conversation, a live-coding technical screen, then an onsite spanning technical depth, a product case study and a deep dive into something you personally owned. Offers have gone out within hours of meeting the right person.

\n

\n

Terms

\n

\n

Base up to USD $300K with founder-level equity on top, calibrated to the individual rather than slotted into a band, so arrive with a number you can defend. Visa sponsorship (new H-1B, TN) and transfers (OPT, H-1B) both supported, with relocation assistance. San Francisco, in the office, five days a week.

\n

\n

Not for you if

\n

\n

You need remote or hybrid. You want the scope defined before you start. Your last two years were advisory rather than hands on keyboard. Or your record shows three consecutive stints under twelve months without a clear upward line through them.

\n

\n

"You've taken your first step into a larger world."

\n

(Obi-Wan Kenobi, A New Hope)

\n

\n

Send your CV with one paragraph on the hardest system you have personally designed, built and shipped, and what broke the first time it met production. The second half of that question is the one they care about.