Fintech One-on-One · Oct 29, 2026 · 33m · 28m saved
With Ilya Fridman, Founder and CEO, Byzfunder
Ilya Fridman, founder and CEO of small-business lender Byzfunder, told Peter Renton on Fintech One-on-One that off-the-shelf bank-statement parsers got loan underwriting right only 76% of the time — so his team wrote down every rule a human underwriter follows and trained nine language models on it instead.
"Can we and the team sit down and write every single rule — right — and train the models on every single rule that we have that a human underwriter does? And we did it." — Ilya Fridman
Ilya started at Morgan Stanley, moved through residential mortgages and merchant cash advance, and launched Byzfunder in 2020, three months before COVID hit.
The full interview is covered here so you can skip it. 33 minutes of audio, 5 minutes of reading.
Here are the 5 takeaways that matter.
This Scuttlebutt Summary is available in full to premium subscribers. If you're not one already, please consider subscribing to keep reading.
Key Takeaways
Byzfunder funds $700 million a year with only 100 employees
Headcount is up about 15% year-over-year while originations are up 75% and unit volume is up more than 150%
Off-the-shelf parsers were only 76% accurate, so Byzfunder built its own
Nine language models, trained on every rule a human underwriter follows, power TraceDataIQ
The new system runs 250 basis points better on delinquency than expected
Not used to cut the default rate — used to approve more businesses at the same charge-off policy
Byzfunder closed a $170 million first securitization, rated by KBRA
The "herculean effort" forced more rigorous, audited processes across the company
He predicts AI produces more entrepreneurs, not fewer
Cheaper expert knowledge lowers the barrier to starting and running a small business
1. Beyond 76% Accuracy
Ilya's core complaint with buying third-party bank-statement parsing: it's built generically for mortgage or accounting use cases, not small-business lending specifically.
"I'm not in the 76% business," he says of off-the-shelf parsing accuracy
"I'm not in the 76% business." — Ilya Fridman
Rather than keep buying that accuracy, Byzfunder's team wrote down every rule a human underwriter follows — reading bank statements, checking background reports, flagging a tax return whose name doesn't match the application — and trained roughly nine language models from different providers, including Gemini and Anthropic models, against those rules. They tested tens of thousands of files with human verification along the way, a project Ilya credits to a jump in base model quality roughly 6-9 months before the interview that finally caught the nuance earlier attempts missed. The result, TraceDataIQ, launched in September.
