A small team built Spring's lending stack, from the bank partnership to the products on top, with AI working in every layer. Here is the stack, bottom to top, and why we are hiring engineers to scale it.
Last edited 5 min read
Lending to small businesses is an old industry that few engineers ever look at closely. Most of the work is reading: bank statements, remittance files, lien filings, retailer portals, email threads. For decades people did that reading, so small businesses got their capital slowly, expensively, or not at all.
Models can now do much of that reading, and that changes what a small engineering team can build. Spring is a lender that builds its own stack, from the bank partnership at the bottom to the products on top, with AI working in every layer. We have funded more than $200 million to over 200 businesses on it. This post walks the stack from the ground up, and ends with why we are hiring.
Our accounts sit at a sponsor bank, and the money we advance comes from a credit facility with a lending partner. When a business joins, we open a dedicated account number for each retailer it sells to. That account is the remit-to on the Notice of Assignment, so an incoming payment already tells us whose it is.
ACH, wire, RTP and FedNow, in both directions. Advances go out when they are approved, repayments come back as scheduled debits or as retailer payments, and a penny test proves every bank account before real money moves. An AI review flags payouts that look wrong before they are released.
Every dollar that moves is posted to one ledger: advances, fees, repayments, deductions. That ledger is what a business sees in its account, what our team works from, and what we report to our lending partner every night. Automated checks, some of them AI, look for anything that does not add up before a report goes out.
We read bank feeds, accounting systems, sales channels and the inbox where remittances arrive, and AI sorts each transaction into what it actually is. Retailer portals are optional. If a business connects them, we pull orders, invoices and payments ourselves; if it does not, it uploads the same documents.
Identity, fraud, liens and credit are checked against the same kinds of data banks use. An AI agent then drafts the credit memo from live data: cash flow, how much depends on each customer, and what the business already owes. For our smallest product the agent also proposes the terms. For everything else an underwriter decides, and every decision keeps the evidence behind it.
Each retailer payment has to be matched to the invoices it settles, every short payment traced to the deduction behind it, and the balance returned to the business. AI reads the remittance details, including the ACH addenda nobody wants to read by hand, and proposes the matches. Debits run on schedule, and collections follows up when a payment runs late.
Invoice factoring, PO financing, Demand Plan and Bridge Funding are thin layers on the same stack, which is how a small team runs four products. The AI CFO our customers talk to reads the same data and uses the same tools as the layers underneath it.
Why we build every layer
When one team owns the whole path, it can fix the slowest step itself. A payment that lands in a dedicated account is matched before anyone looks at it. A decision made from a live bank feed does not wait for statements. An advance can go out over real-time rails the hour it is approved. Most of our speed comes from no step waiting on another company.
It also makes risk easier to see. We know how a retailer pays because we service its invoices. We know an account is real because we moved a penny into it. Underwriting and servicing share one ledger, so each improves the other.
What AI changes about the work
In every layer, the expensive work is reading and deciding. When models get better at reading, every layer improves at once: memos get sharper, matching gets more complete, reviews get faster.
So engineering here is less about building forms for an operations team and more about building the agents that do the work, the evaluations that tell us when they are right, and the guardrails for when they are not. Money moves on these decisions, so every agent's output is checked, traced, and reviewable by a person.
Where the stack is under load
A small team built all of this. The number of businesses we fund has grown quickly, and our newer products reach businesses beyond the consumer brands we started with. Every layer in the diagram now carries more than it was built for: more accounts, more payments to match, more memos to draft, more data to connect.
Some of what comes next:
- Matching and servicing that keep up with a growing ledger without more people reviewing by hand.
- Underwriting agents we can measure, and trust with larger decisions.
- Connections to more data sources, for more kinds of businesses.
- Money movement that stays correct as volume grows.
You would work directly with the founders, decide what gets built, and own it from design to production.