cited memo from the file7-day trial · no cardSOC 2 Type II · published evals
SecureLend
Live AI underwriting

Turn a deal file into a cited underwriting memo.

Upload a borrower package, pitch deck, or data room. SecureLend extracts the facts, flags risks, applies your rubric, and drafts the memo. You review the work. You do not redo it.

7-day free trial · no card to startSee how it works

Product preview

Example underwriting fileAcme Manufacturing$2.0M revolving facility · LTM 2026

Agent workflow

  1. Package indexed14 files · 286 pages
  2. Financials spreadFY23–LTM normalized
  3. Risks challenged3 exceptions surfaced
  4. Memo drafted18 citations linked
Review requiredPolicy exceptions surfaced

Customer concentration, covenant headroom, and a filing mismatch are routed to the underwriter for judgment.

3exceptions to review

Credit memo · draft · Human approval required

Commercial lending 1 of 3

Controlled by your team

Your models. Your data. Your controls.

Put the harness inside the operating model you already trust. Every material finding stays connected to the source, policy, and human decision behind it.

Your approved models

Use SecureLend defaults, route steps by task, or connect supported model providers with encrypted credentials.

Your institutional data

Combine deal files with the systems and data sources your team already relies on.

Your underwriting policy

Apply your rubric, escalation rules, and approval gates instead of a generic scoring template.

Decision recordReview required

Recommendation

Approve with conditions

Sources18 citations

PolicyCommercial v4.2

Exceptions3 open

  • Financial spreading reconciled to source pages
  • Policy exceptions routed to the assigned reviewer
  • Final decision reserved for an authorized human

Material actions are logged with their evidence and review state. Customer-provided third-party API credentials are encrypted at rest.

1 / 2

Bring your own model

Route each job to the model you approve.

Use SecureLend-managed routes or connect a tenant-owned OpenRouter key, then apply model policy by workload, quality tier, and cost. The evidence workflow stays consistent while the selected model can change.

SecureLend-managed routes

  • OpenAIManaged
  • AnthropicManaged
  • GeminiManaged
  • DeepSeekManaged

OpenRouter model catalog

  • Z.AI · GLMOpenRouter
  • KimiOpenRouter
  • MiniMaxOpenRouter
  • QwenOpenRouter
  • xAIOpenRouter
  • MistralOpenRouter

OpenRouter inference & self-hosted

  • Together AIOpenRouter
  • FireworksOpenRouter
  • OllamaPlanned

Managed routes are built into SecureLend. OpenRouter entries use the tenant’s OpenRouter account and its live catalog. Ollama is a self-hosted roadmap item: the current fixed OpenRouter proxy does not connect to local Ollama endpoints.

Who it is for

§ 03

Underwriting venture capital, private equity, mortgages, business loans, or public equities. The harness and the analyst cover the file.

VC

Venture capital

IC memo from the deck and CIM.

PE

Private equity

Diligence file. Same harness.

RE

Mortgages

Package in. Cited memo out.

LOAN

Business loans

Credit file, ratios, you decide.

EQ

Public equities

Filings in. Memo out.

Public benches

Four separate tracks. Never a blended SecureLend score.

Credit files, origination, bank statements, and IB file shape are different jobs. We publish each result on its own — we do not average them into a SecureLend number.

Pilot measurements. Not official leaderboard entries.

UWBench

Cited memos on credit files. Our MCP chat vs general coding harnesses, by lane.

90.3 commercial credit

Listed SME 91.3 · raw documents 91.0. Lanes are not averaged. Not official UWBench scores.

uwbench/uwbench

LOAB

Origination tasks. Live MCP vs the official LOAB scorer.

24/24 outcome

Full-rubric 20/24. Not 24/24 on the full rubric. Claude Opus 4.6 20/23 outcome.

shubchat/loab

MortarBench

Bank-statement transaction QA. Mean of 3 trials, 90 items, n=270.

95.5 F1

EM 90.0. Clears Gemini 3.1 Pro (80.8 / 77.1) and + CRIT (83.6 / 80.5).

mtoles/MortarBench

Self-reported · unpublished Harbor fill · not an official paper leaderboard

BankerToolBench

Harbor grades file shape (tabs/layouts). Two results, reported separately: the unpublished OpenCode×grok-4.6 Harbor board, then a later fixed 5-task sample.

OpenCode×grok-4.6 Harbor board

100/100 Mean 0.6843

2026-09-13 · unpublished Harbor fill · median 0.733 · range 0.195–0.980 · all eval_pct=100

OpenCode×grok-4.6 Harbor board / quality sample. Not an MCP n=100 run.

Subsequent fixed 5-task sample

5/5 beat

SecureLend MCP beat OpenCode×grok-4.6 on every pack in the subsequent fixed 5-task sample.

Pack mean ~0.821 vs ~0.654

Subsequent fixed 5-task sample. SecureLend MCP vs OpenCode×grok-4.6. Not an MCP n=100 run.
PackIDOpenCode×grok-4.6SecureLend MCP
AMZN41e130fe0.74880.9064
Okta69fcdefe0.67490.7262
Hyperscalers60ed36b00.9160.9262
Energy5f1e6f380.3810.6476
OTA922356490.54980.8961

Self-reported. Unpublished OpenCode×grok-4.6 Harbor fill — 100/100 mean 0.6843 is the OpenCode×grok-4.6 Harbor board, not an MCP n=100 run. The 5/5 beat is SecureLend MCP vs OpenCode×grok-4.6 on a later fixed five-pack sample. Not an official paper leaderboard.

Handshake-AI-Research/bankertoolbench · laude-institute/harbor

Try it free

§ 04

Seven days of Heavy. No card. A real file, a cited memo, you still sign.

Start 7-day trialBook a walkthrough

What it costs

§ 05

Analyst

$1,500/mo

One person, full product. Drafts and challenges the file. You still sign.

Start 7-day trial

Heavy

$3,000/mo

Same product, more files at once. Trial is Heavy for 7 days, no card.

Start 7-day trial

Extra person on the same plan: 50% off.

Pay for the analyst. Pay for our books only if we hold the loan. Their nCino or spreadsheet: $0 book fee.

Trust

§ 06

SOC 2 Type IISOC 2 Type IITrust CenterUWBenchpublished evals

Works with the model you already trust

  • OpenAI
  • Anthropic
  • Gemini

Human approval on credit. Nothing invented when a check fails.

Compare and read

§ 07