Commercial lending
Our own lending business. Borrower packages are read and cross-checked, and lenders are matched to each deal, with every figure tied to its page.
Auditable AI decisions · the guardrail for AI agents
Every number, traced to its page.
Habena is the guardrail for AI work where one wrong number is costly. A model reads every page. Code checks every citation and holds the veto on hard rules. People make the final call, with the source page already open.
Habena
Founded by Blake Janover
U.S. citizens · U.S.-based
In production in our own lending business
The problem
Agents need guardrails.
The Department of War has almost 1.7 million people on GenAI.mil, and they have built more than 100,000 agents.1 Wherever one wrong number is costly, from clearing a supplier’s foreign ownership to financing a defense manufacturer, those agents need guardrails. It’s the same problem across agencies that lend, license or buy, and in banks.
01
Thousand-page packets
Filings, manuals, applications and scans, in every format, still read and cross-checked by hand.
02
Sources that disagree
The same company, owner or number written four ways across four documents.
03
No proof of source
Language models read well, but they can’t prove where a number came from. In one published test on company filings, a leading model with retrieval got 81% of questions wrong or declined to answer.2
04
A different answer each run
Even when set to be consistent, a model’s accuracy can swing by up to 15% from one run to the next.3
How it works
The model reads. Code verifies. People decide.
Read
A language model reads every page and returns each figure with its page and the exact words behind it.
Verify
Code checks the model’s work: the quoted words must be on that page, and the number must be inside them. Then fixed checks run across the documents.
Decide
Code holds the veto on hard rules. Anything ambiguous goes to a person with the source page open. Nothing gets guessed.
- Figure
- Largest foreign beneficial owner: Example Holdings Ltd. · 31.4%
- Source
disclosure.pdf · page 14- Quote check
- Quoted words found on page 14; the number is inside them
- Cross-check
- Owner listed in 3 of 4 filings, missing from one
- Hard rule
- In code: foreign voting interest of 25% or more
- ExceptionSent to an analyst with page 14 open.
Every job goes to the model that fits it. Frontier or open-weight, swap any model without touching a rule. Every model call is logged and costed, and anything unresolved goes to a person.
Compared with the alternatives
Reads like a model. Holds like code.
| Approach | Reads messy documents | Proves each number | Same answer every time |
|---|---|---|---|
| Manual review | Yes, slowly | Sometimes | No |
| Rules-only software | No | Yes | Yes |
| Chat assistants and agents | Yes | No | No |
| Habena | Yes | Yes | Yes, for every rule in code |
Manual review is careful but slow. Rules-only software is consistent but breaks on messy documents. Chat assistants read anything but can’t prove an answer. Habena puts the models to a new use: they read, code verifies, and every answer is cited. It runs in production today in our own commercial lending business.
In production, and what it can be set up for
One harness. Your documents, your rules.
Supplier and foreign-ownership vetting
Ownership filings and registries checked against your foreign ownership, control and influence thresholds. An owner listed in one filing and missing from another gets flagged.
Defense-industrial and federal credit
Applications, tax returns, financials and appraisals reconciled and checked against credit and covenant rules.
Sustainment and parts
Technical manuals, scanned drawings and maintenance logs matched to National Stock Numbers.
For older systems, it reads from and writes back to SQL and Oracle databases in place.
Environments
Your environment. The models you already use.
| Environment | How it runs | Status |
|---|---|---|
| Your cloud account | Installed in your account, connected to the model endpoints your office has approved | Designed for |
| Your own servers | Installed on your hardware, with open-weight models such as Llama or Mistral | Designed for |
| Disconnected networks | Same as your own servers, with no outside calls | Designed for |
| Our cloud | Our own commercial lending business | In production |
We have not deployed on a government network yet, and this table will say so until we have. For government work we never host your data. Model calls go straight from your environment to your endpoint, so your data-retention terms carry over. Until the software is inside your boundary, we work only from blank templates and fictitious records.
Business model
Price follows scope. No per-seat fees.
Fixed price, paid by milestone. When it’s done, it’s yours to run.
01
Audit
We audit one of your workflows with your team.
02
Scope
The price follows the scope: how many document types, rules and systems we connect.
03
Build
Built at a fixed price, paid by milestone on acceptance.
04
Hand off
Installed in your cloud or on your servers, connected to your models. Yours to run, with no subscription.
Questions contracting officers ask
Straight answers.
How does a program office buy this?
Most often through a prototype agreement under 10 U.S.C. § 4022, reached through a Commercial Solutions Opening or the Tradewinds Solutions Marketplace. If the prototype is awarded competitively and completed successfully, follow-on production can be awarded without further competition (§ 4022(f)). Email Blake and we’ll work out the path with your contracting officer.
What does it need from us?
Two things: a model endpoint your office has approved, and access to the documents. Risk is managed in code: any model can be swapped, every output is checked, and people make the final call.
Is there a cost-share requirement?
We expect Habena to qualify as a nontraditional defense contractor (10 U.S.C. § 3014). That would let a prototype agreement meet the statute’s cost-sharing condition without a one-third cost share (§ 4022(d)(1)(A)).
Have you sold to the government before?
No. The harness runs in production in our own commercial lending business, and our companies have built their own software for nearly twenty years. We would rather tell you that here than have you find out on a call.
Do you host our data?
No. Habena installs in your cloud account or on your own servers, and model calls go straight from your environment to your endpoint. Until it’s inside your boundary, we work only from blank templates and fictitious records.
Which AI models does it use?
The ones you already use: a frontier model your office has approved, or an open-weight model such as Llama or Mistral on your own hardware. Each job goes to the model that fits it, and swapping a model doesn’t touch a rule. Every call is logged and costed.
Who writes the code?
A small team of U.S. citizens in the United States. No offshore contractors.
Who owns what at the end?
When it’s done, it’s yours to run: a permanent, royalty-free license, plus full rights to the rules and connectors built for you. The core harness stays ours, so we can keep improving it. No subscription.
What does it cost?
The price follows the scope: how many document types, rules and systems we connect. We audit the workflow with you first, then quote a fixed price paid by milestone. No subscription, no per-seat fees.
Team
Why I’m building this.
I’ve spent my career building companies and sitting on public-company boards, and my companies have built their own software for nearly twenty years. I’ve never sold to the government. I’m starting Habena because I love this country, and too much of the work that keeps it safe still comes down to someone reading stacks of paper by hand.
We’re a small American team. Our code is written here, by U.S. citizens. We’re new to federal work and we expect to earn trust one file at a time. If our software can’t show you the page behind an answer, it shouldn’t be making the call.
— Blake Janover, Founder
The team
Blake Janover
Founder
3U.S. public listings as founder, CEO or director: Janover Inc. (Nasdaq: JNVR), Soulpower Acquisition Corp. (NYSE: SOUL), Deep Fission (Nasdaq: FISN).
- Founded Janover Inc., a commercial real estate finance marketplace, and took it public in a 2023 Nasdaq IPO (JNVR). Co-architected its transition to DeFi Development Corp. (Nasdaq: DFDV), where he serves as a director.
- Director, Deep Fission, selected for the U.S. Department of Energy Reactor Pilot Program. Chairs its compensation committee.
- NATSEC Fellow, National War College Alumni Association.
- Harvard Business School, OPM 60. YPO Palm Beach board. Lecturer and mentor, Zell Entrepreneurship Program, Reichman University.
Steve Schwartz
Chief Architect
- SVP of Innovation, Janover Inc. The company’s third-ranking executive, after the CEO and CFO. Led its technology through the 2023 Nasdaq IPO.
- Built lending software that processed billions of dollars in loan requests, and has been a key architect of its AI-first systems and automations since ChatGPT arrived.
- Chief Product Officer, J2 Labs. Nearly 30 years as a developer and entrepreneur.
Contact
Send us your hardest stack of paper.
Tell us the decision and the documents behind it. Blake reads every message.
blake@blakejanover.comWhat to include
- The decision you need to make, and how often it comes up.
- The documents behind it: what kinds, and roughly how many pages.
- The rules it has to follow.