What we did over the summer — built our own platform using A.I.
Being inside a scaling company can be a difficult. If you have ever read the Shoe Dog by Phil Knight the story of Blue Ribbon and Nike, there are a lot of similarities we can draw as a growing company juggling lots of things in the early years.
Most of what we do at Gallantree is self described as being hand-built. We review each opportunity by build out financial analysis, conduct quantitative and qualitative modeling, draw up customised Investment Memorandum’s and compile data rooms for ourselves, financing partners and syndicates. It’s really no different to that of most financiers, however we are a small team with a limited budget.
Coming from a fintech background, I started coding an API to a MongoDB as a basic challenge to myself to keep the ‘skills up’. It was fairly rudimentary but was built in a way that it one day could be the backbone of what we do as I considered the overall architecture of the platform.
This summer in Queensland it was incredibly wet, therefore with limited time outside, I read books (Black Edge, Shoe Dog) and began tinkering with the API with a simple goal of being able to move one of our spreadsheets into a platform where the team can login and get a coupon rate. I quickly learned that Chat GPT 3.5 and Tabnine — two AI tools would speed me up significantly. I can safely say over a 14 day period, I did more with those tools than a small team would in 6 months.
What did we build?
In essence, we ended up building our world in the platform. Traditionally we’ve used Box for storage, Hubspot for CRM, Creditorwatch for credit checking, RapidID for KYC, Slack for communications (internal) and excel and docs for the heavy lifting.
const express = require("express");
const session = require('express-session')
const cors = require("cors");
const dotenv = require('dotenv');
const app = express();
const axios = require("axios");
const bodyParser = require("body-parser");
const chatGPT = require("chatgpt");
// These libraries make things faster!All of these tools are now in our platform and interacting with each other. I started building JWT authentication and authorization, then I moved onto constructing users and organisations. Once I got there I wanted to enable opportunities for property credit therefore I begun building references to products, fees, coupons, locations, asset types, borrower pedigree / verification and funders.
For some of the technology people out there, we built a nodejs API and Dashboard with a combo of libraries and SDKs linked to a MongoDB and hosted on a combination of AWS and Heroku.
Here is a quick example of what I mean when using Chat GPT for coding.
Of course, you need to know what it’s doing, but it can really speed you up with mostly error free, curly bracket error free code.
What does our platform do?
I wanted it to solve at least four problems and integrate them using off the shelf AI tools to automate:
- Create a mini CRM to replace Hubspot eventually whereby we can add contacts and companies and integrate with our affiliates and our website.
- Enable automatic Organisation and client verification — we like to make sure the people we are lending to are trustworthy and dont have skeletons in the closets. AML CTF requirements is like step 1, we wanted to enable step 10 and look at a range of other data from courts, media, ASIC and other platforms to complete a full picture. AI has been implemented to focus on ‘hot areas’ against a tonne of data coming back to point out what we should be cautious on.
- Enable dynamic credit analysis per opportunity and autonomously using advanced learning — For property opportunities, we look at a range of qualitative and quantitative data from market data and asset concentration in the area (Gold Coast needs more resi skyscrapers) to the economics and fundamentals of the opportunity like min/max LVR, borrow amount against credit score etc. This not only produces a coupon rate, but considers how competitive we should be against metro vs regional locations, borrower quality, asset evaluation, market demand etc.
- Produce at scale lending documentation and IC documentation in seconds — The best lenders produce term sheets anywhere from 2 hours to 3days and loan documents from 2 days to 2 weeks. For those lenders that follow a credit committee pathway with pre-approval and approval to disburse funds, the documentation build can take 3–4 best case, 14–16 days for the poor ones. Our challenge was, in a series of steps, being: 1) create indicative term sheets and legal documentation within seconds against lending parameters, and 2) enable start to finish disbursement within 36 hours that is compliant and accurate. The documents use data from the database to construct ontop of a template word doc and using AI, validates the text and language used. The Investment Memorandums for credit committee or syndicate partners offer a ‘traffic light’ system and validation for validation on whether we should lend.
On point 4, we are naturally only as good as the capital we can source, however that is another blog coming.
How does it look?
It’s fairly basic, but it does the job. As mentioned, currently this is an internal tool, not an external. In the image below, admin staff have added an organisation and enable it from pending to active by verification of points 1 and 2 above. They are creating a Property Finance opportunity now and entering the product, funding source, amount, coupon details etc.
And here is some functionality:
Where to scale from here?
For Christmas, each of us received a book called the Black Edge by Sheelah Kolhatkar on the rise of Steve Cohen. For most, they may not have heard of him but they may have heard of Axe Capital and Billions where Steve is the real version of Axe.
While it’s a great story, one thing resonated and it’s spoke about in Jeremy Irons character in Margin call as well.
There are three ways to make money in this world, 1) be first, 2) be smarter, or 3) cheat — and we don’t cheat.
Be smarter is the one for us. We’ve got a white board full of ideas, concepts and plug in not only property data but a world of data points amplifying our understanding or credit, borrowers, trends or markets.
Building our own platform vs buying an off the shelf Loan Management Solution (LMS) and Customer Relationship Management Solution (CRM) gives us flexibility to code a concept, an integration or even client integration in an afternoon, not in 2 months.
Whats the cost of the platform to date?
So far, it’s $350.00 in server cost and around 150 hours of coding over around 12,000 lines of code. The cost will increase as more API’s are turned on, more servers are required and more security is implemented.
What’s next?
We’ve pretty much got what we wanted from Property Credit side, so now we want to consume data for Corporate Credit for specific verticals we want to go after. Agriculture, resources, financial services are some areas, but we’ve even found niches available in specific areas where we can do dynamic verification on credit assessments — which we believe is the next step.
Additionally, one of Bloomberg’s greatest strengths is network connections which is basically an internal LinkedIn of the financial services world.
There are truly not enough hours in the day!
