Pylon Lending
Pylon is “the first AI-native infrastructure platform to deliver autonomous mortgages at scale,” handling “everything from application to capital markets settlement” (home) — betting that “mortgages are the last major financial product that are not programmable” and turning origination into “a single API” (about). The engineering core: compile dense underwriting guidelines into executable DSLs, orchestrate the multi-day application-to-settlement pipeline on Temporal, and rep-and-warrant every loan into the capital markets.
Vitals: founded ~2022 · ~$45M raised (Seed→A) · ~40 people · Palo Alto (eng) + NYC (GTM).
Business context — team, funding, results, products
- The team “primarily comes from Stripe and Better” (introducing); HQ Palo Alto (engineering) + New York (GTM/Product) (careers, Ashby); ~40 people (Paraform tracker).
- $45M raised (about) — seed ~$8M led by Conversion Capital (Dec 2022, FinTech Global); backers Conversion Capital, Peter Thiel, QED, Citi, Allegis Capital, Fifth Wall (introducing), plus angels incl. founders of Ramp, Mercury, Blend, DoorDash, Wealthfront and Naval Ravikant (about).
- Citi connected its “mortgage trading desk to the Pylon platform alongside a strategic minority ownership investment” (introducing).
- Claimed results: 74% lower cost to originate vs. the Freddie Mac 2024 study, 75–200bps better pricing, ~102bps more per loan, 2× MoM revenue growth (home, careers).
- Five composable products on the API — Decisioning, Capital, Command Center, Elements, Compliance (introducing) — sold to brokers, fintechs, lenders, and banks (home).
- The build order is the bet: Pylon re-wired the stack “from capital markets to initial application (in that order)” (introducing). Owning the takeout first — it “reps & warrants every loan into the capital markets” (fintechs) — is what lets it collapse the rest of the chain into an API instead of reselling someone else’s pipeline.
The heavy lifting
Section titled “The heavy lifting”- Underwriting compiled to a tested DSL, not staffed. Dense regulatory guidelines are encoded — with AI as draft-generator — into executable DSL rules built “side-by-side with mortgage experts”; a mis-encoded rule costs “tens of thousands of dollars on a single loan” because Pylon reps & warrants the loan (Underwriting JD, fintechs).
- Temporal orchestrates every multi-day origination workflow. A mortgage is multi-party and failure-prone across days; Temporal (named in every eng JD) gives durable, replayable workflows over Postgres state, with humans as exception handlers (API JD, Infra JD).
- One GraphQL API encodes a path-dependent domain. The contract models mortgage branching and path-dependence natively rather than papering over it, and is treated as the product customers build on — versioned, event-driven, surfacing “what happened, why, and what you can do next” (API JD).
A single-language TypeScript shop: TS everywhere, GraphQL as the contract, NestJS services, Postgres for state, and Temporal.io orchestrating every long-running mortgage workflow — on AWS. Every row below is named in a first-party JD or shown on the product site.
| Layer | Choice | Evidence |
|---|---|---|
| Backend language | TypeScript (everywhere) | every eng JD (API, Infra, Underwriting, Integrations, Fullstack) |
| API contract | GraphQL — “this is the product” | API JD; createBorrower mutation w/ query-complexity limits on home |
| Backend framework | NestJS | API, Underwriting, Integrations, Fullstack |
| Frontend | React | Fullstack JD |
| Primary datastore | PostgreSQL | every eng JD |
| Workflow orchestration | Temporal.io | every eng JD |
| Architecture pattern | event-driven | API, Infra |
| Cloud | AWS (“our cloud home”) | Infra JD |
| Code review / dev flow | Graphite (stacked PRs) | Infra JD |
| Observability | Honeycomb | Infra JD |
| Dev environments | GitHub Codespaces | Infra JD |
| Underwriting rules | custom DSLs + AI (guideline → executable logic) | Underwriting JD |
| External integrations | credit bureaus, title, insurance, doc services via REST / SOAP / file | Integrations JD |
| Agent access | MCP-native infrastructure | home, fintechs, developers |
| AI in product + dev | ”AI-driven development tooling and agentic infrastructure” | every eng JD |
| Marketing site | Next.js, HubSpot, GTM | network trace (careers) |
Hard problems
Section titled “Hard problems”The parts an engineer would lose sleep over. Public signal is cited (verified); likely approach is labeled speculation — best-practice fill-in, hedged.
| Problem | Why it’s hard | Public signal | Likely approach (speculative) |
|---|---|---|---|
| DSL rule correctness | A single mis-encoded guideline is Pylon’s own loss, not the borrower’s | ”An incorrectly modeled rule can cost the company tens of thousands of dollars on a single loan”; Pylon “reps & warrants every loan” (jd-uw, fintechs) | Likely a golden-file/snapshot test suite over historical loans plus property-style tests, gating every DSL change in CI |
| AI-compiled guidelines | An LLM translating dense regulatory English into executable logic can be subtly, expensively wrong | ”translate dense regulatory guidelines into executable logic … including AI … compiling English into a system that makes six-figure decisions,” built “side-by-side with mortgage experts” (jd-uw) | Probably human-in-the-loop review of generated rules with the DSL as the audited artifact, AI as draft-generator not decider |
| Durable multi-day workflows | A mortgage is a multi-day, multi-party flow where a dropped step or bad replay loses real money | Temporal in every eng JD; Foundation keeps “highly available systems” that “process millions of dollars in mortgage transactions” (jd-infra, jd-api) | Likely deterministic Temporal workflows with idempotent activities and saga-style compensation for external order-outs |
| Per-file compliance + auditability | Every automated decision must be explainable and TRID-compliant on demand for regulators and capital-markets buyers | Pylon “encodes compliance into every file”; settlement enforces TRID timing and a zero-variance fee ledger (fintechs, home) | Probably an immutable per-loan decision/event log tying each outcome to the exact DSL rule version that produced it |
Likely internals
Section titled “Likely internals”The infrastructure Pylon doesn’t name publicly, inferred from the stack it does:
| Component | Likely choice | Basis |
|---|---|---|
| Reasoning LLM | a frontier model behind a provider abstraction, used for guideline compilation + doc understanding | ”including AI” in the underwriting DSL (Underwriting JD); no model named |
| Underwriting DSL form | a TypeScript-embedded DSL or a rules engine, not a standalone language | ”custom DSLs for rule encoding” (Underwriting JD); standalone vs embedded unstated |
| Document AI / OCR | a managed doc-extraction service or in-house ML for paystubs, bank statements, title | Intake “verifies income and assets” (home); extraction is implied, not named |
| Guideline retrieval | embeddings + vector index over investor guidelines | ”mapping file to guidelines” (home); retrieval over dense rulebooks is the natural fit |
| Event backbone | AWS-native (SNS/SQS/EventBridge) or Kafka feeding the event-driven API | ”event-driven architecture” on AWS (API, Infra) |
| Container orchestration | ECS or EKS on AWS | AWS confirmed (Infra JD); orchestrator not stated |
| Auth / identity | a managed IdP for platform + embedded borrower flows | enterprise/regulated buyers; white-label Elements; no vendor named |
Architecture
Section titled “Architecture”One GraphQL API fronts a five-stage, Temporal-orchestrated pipeline — Intake → Structure → Process → Underwrite → Settle — that reaches out to external mortgage rails for data and lands every loan in the capital markets. Customers reach it three ways: browser, API, or agents over MCP.
Mermaid source
flowchart LR classDef client fill:#eef2f8,stroke:#94a3b8,stroke-width:1.5px,color:#0f172a; classDef api fill:#e8f1fd,stroke:#2563eb,stroke-width:1.5px,color:#0f172a; classDef stage fill:#eef0fe,stroke:#6366f1,stroke-width:1.5px,color:#0f172a; classDef engine fill:#eafbf1,stroke:#16a34a,stroke-width:1.5px,color:#0f172a; classDef ext fill:#fff7ed,stroke:#d97706,stroke-width:1.5px,color:#0f172a; classDef human fill:#fdecec,stroke:#e0564f,stroke-width:1.5px,color:#0f172a;
Cons("Originators<br/>brokers · fintechs · lenders · banks"):::client Surf("Integration surface<br/>Browser · API · Agents (MCP)"):::client
API("GraphQL API · the product<br/>TypeScript · NestJS · event-driven"):::api
subgraph Pipe["Origination pipeline · orchestrated by Temporal.io"] direction LR Intake("Intake<br/>eligibility · income/asset<br/>verify · preapproval"):::stage Struct("Structure<br/>pricing · rate-lock<br/>best-option"):::stage Proc("Process<br/>order-outs · disclosures<br/>conditions"):::stage UW("Underwrite<br/>DSL + AI rules<br/>AUS · compliance · fraud"):::engine Settle("Settle<br/>closing disclosure · eSign<br/>notary · recording"):::stage Intake --> Struct --> Proc --> UW --> Settle end
PG[("PostgreSQL<br/>loan state")]:::api
subgraph Ext["External rails · REST / SOAP / file"] direction TB Bureaus("Credit bureaus"):::ext Title("Title · insurance"):::ext Docs("Document services"):::ext end
Cap("Capital-markets takeout<br/>direct Wall St · Citi trading desk<br/>rep & warrant every loan"):::ext Ovr("Technical Oversight<br/>licensed mortgage ops · exceptions"):::human
Cons --> Surf --> API --> Pipe Ext --> Proc Ext --> UW Settle --> Cap Pipe -. "state in" .-> PG Pipe -. "exceptions escalate to" .-> OvrThe API encodes the domain, not a wizard. Pylon models mortgage as “path-dependent” and “nonlinear” — “the choices you make early in a loan constrain what’s possible later … different borrower situations branch into wildly different flows” — and bets that the API should “encode choice, branching, and path-dependence natively” rather than “paper over that complexity” (API JD). They are moving it toward event-driven: an API that “doesn’t just respond to requests — it tells you what happened, why, and what you can do next” (API JD).
Underwriting is a compiler, not a checklist. The underwriting team “takes human judgment out of mortgage origination and replaces it with systems that are faster, more consistent, and more accurate” (Underwriting JD). The mechanism: “Encode natural language rules into code … with DSLs and novel techniques — including AI — to translate dense regulatory guidelines into executable logic … compiling English into a system that makes six-figure decisions,” built “side-by-side with mortgage experts.”
Integrations are the connective tissue. A dedicated team plugs Pylon into “credit bureaus, title companies, insurance providers, document services” — “mortgage touches everything” — across “REST, SOAP, file-based — the full spectrum” (Integrations JD). The pipeline’s order-outs, disclosures, AUS, and fraud checks all ride these connectors, and Settle hands the finished loan to the capital-markets takeout (the Citi trading desk and direct Wall Street access).
Team & process
Section titled “Team & process”Small, senior, and pointedly from outside mortgage: “We don’t come from the mortgage industry. We came in from the outside, got obsessed with the problem” (careers). 30% are former founders; the engineering bio reads “Many ex-Stripes” (careers, API JD). All Palo Alto / hybrid at $130–220K + equity (Ashby).
| Role | Person | Background |
|---|---|---|
| CEO | Trent Hedge | ex-founder, Atmos (about, introducing) |
| CTO & Head of Pylon Labs | Josh Kuhn | Stripe, Theorem (about) |
| VP Engineering | Yves Bourelle | Stripe, Box (about) |
Engineering is organized by mortgage subsystem — API (the GraphQL surface), Integrations (external rails), Underwriting (the DSL/AI engine), Foundation (infra), Customer Success (fullstack), SRE — and engineers “own entire systems, not tickets,” leveraging “AI+ML and operations research” (careers, Ashby). Licensed mortgage pros don’t manufacture loans here: a separate Platform department staffs remote “Technical Oversight” roles (veterans of Better, ICE Mortgage, Mr. Cooper, loanDepot) that supervise the automation and handle exceptions (Ashby, about). The dev loop is AI-native — Graphite stacked PRs, Codespaces, Honeycomb, and “AI-driven development tooling and agentic infrastructure” in every JD (Infra JD).
Sources
Section titled “Sources”Reconstructed from public sources only — no insider information. Crawled 2026-06-07. Claim tiers: verified (stated on a public page, linked) · inferred (reasoned from a cited signal, confidence flagged) · speculative (best-practice fill-in, labeled). Links are live; pages change, so the supporting quote for each claim is kept in this repo’s evidence map (evidence/pylon-lending-evidence-map.md). Note: this is Pylon Lending (pylonlending.com), not the unrelated “Pylon” customer-support SaaS.