Niral J Shah

FinTech / Data / Plain writing

I build things you can inspect.

Lending software. Credit models. Data projects with the working attached.

SOURCE RULE EVIDENCE SHOW THE WORKING. TRACEABLE OUTPUT
  • FUNDO LOANS$1M MRR → 3× multiple at acquisition, 1.5× the loan book
  • CREDITCREST23 modules in 4 suites
  • TESTS11,000+ guard every Creditcrest release
  • PACKAGES0 third-party dependencies
  • AUTOMATIONover $265K saved, approved by PwC
  • OCRaccuracy 10% → 74%
  • REPORTINGdata-to-report 14× faster
  • MANUFACTURINGUS$38,150 saved in 2 months
  • TOOLSfault prediction 65% → 90%
  • LOGISTICS74.45% accurate, built in 15 days
  • PYPIpip install basicanalysis
  • WALKING CHALLENGE10.5M steps in 8 weeks

Selected work

Four suites. One principle.

Creditcrest Technologies

23 modules · 4 suitesOpen to try

Lending infrastructure for Australian credit providers, from the first screen of an application to the last report to the credit bureau. Take a whole suite or any one of its modules: every figure comes with the evidence behind it.

Try every product (opens in a new tab) · Creditcrest Technologies is not a credit provider, is not a credit assistance provider, and does not hold an Australian credit licence. Its software produces evidence; the licensee makes the decision.

Explore Creditcrest (opens in a new tab)

Hi, I’m Niral

I turn messy financial data into numbers you can check.

I trained as an electronics engineer and have spent the decade since where data meets money. As lead data scientist at Fundo Loans, I helped a small team take an Australian lender from $1M MRR to a 3× multiple at acquisition, with 1.5× the loan book, building its data architecture, automated decisioning and bank-statement analysis from scratch.

Before that: credit and fraud models for lenders, insurance risk at PD Insurance, yield engineering at Maxim Integrated (Now Analog Devices), and a master’s in Data Science at Monash. On the side, I’m building Creditcrest, lending software for Australian lenders, on one rule: every number comes with the working that produced it.

Off the clock: table tennis, stargazing and listening to big band.

Worked withFundo LoansCash DirectTrimbleMonash UniversityPD InsuranceMaxim Integrated (Now Analog Devices)

The full story & CV
Building Creditcrest · Parramatta, NSW

The average of everyone is a reasonable place to start. It is almost never the real answer for you.

The model supplies the average. The working, and the judgement, are still yours.

From electronics to evidence.

  1. 2011–15 B.E. Electronics & Communication Vishwakarma Government Engineering College
  2. 2016–18 Yield Engineer Maxim Integrated (Now Analog Devices)
  3. 2018–19 Data Analyst PD Insurance
  4. 2019–20 Research Assistant Monash University, SoDa Labs
  5. 2019–21 Data Scientist, self-employed Dusk Mobile · Trimble · Cash Direct
  6. 2020 Master of Data Science Monash University
  7. 2021–now Lead Data Scientist Fundo Loans
  8. Now Building on the side Creditcrest Technologies

Now

What I’m working on.

Updated

Creditcrest

  1. Founding lendersTen places for the first lenders, at 40% off for three months, until 30 June 2027.
  2. The Statement APIA bank statement read from a lender’s own program: a sandbox key in one call, and the accuracy published.
  3. Research for lendersA monthly brief on small lending in Australia, and what the year’s prices did to each kind of household.

AI & data

  1. Ready for AI agentsAn llms.txt for Creditcrest’s site and API, and an OpenAPI 3.1 file for code generators.
  2. AI and Data AdvisoryA lender’s data, models and AI governance, worked on with them by people.
  3. LLMs and AI for your daily useMy notes on asking a model for the specific, getting past the average, and checking its working.

Projects

Things I’ve built, measured and shipped.

All 9 projects
  • 2019

    Stock price analyser

    The daily sentiment of the news about Facebook, averaged and read against its share price from June to July 2019.

  • 2019

    ELO-rating model for loan default

    A credit score built the way chess ranks players, from bank statements alone, with a confidence interval that says when to buy a credit report instead.

  • 2019

    basicanalysis (opens in a new tab)

    A small Python package that runs eight supervised-learning methods over a dataset in one go and compares them on eight measures, from accuracy to log loss.

  • 2019

    Fall Relief

    A team of four building targeted exercise and diet plans to reduce falls among people aged 65 and over.

  • 2019

    Document OCR at SoDa Labs

    Research data compiled from scanned documents at Monash University: a digit-recognition network built from scratch at over 90% accuracy, and a dedicated recognition model that took extraction accuracy from 10% to 74%.

  • Next

    The next one is in the works.

    I’m always building something. New projects land here first, and in the newsletter.

What I know well

FinTech, data, and the models that read both.

  • Lending & credit

    Automated decisioning and bank-statement analysis at Fundo Loans. On the side, Creditcrest’s four suites for Australian lenders, each figure showing the evidence behind it.

    • Decisioning
    • Serviceability
    • Credit reporting
    • Best interests duty
  • Data science & ML

    Credit scores and fraud models for lenders, an ELO-style credit score, and an OCR model that took extraction from 10% to 74%: each measured against the question it was built for.

    • Supervised learning
    • Credit scoring
    • OCR
    • Forecasting
  • Data pipelines

    Data architecture built from scratch at Fundo Loans, monthly reporting automated to run 14 times faster at PD Insurance, and analysis reports automated across Maxim’s sites.

    • Data architecture
    • Automation
    • SQL
    • Reporting
  • AI & language models

    Language models asked for the specific and checked for their working: context first, sources quoted, assumptions listed. And Creditcrest’s API, ready for AI agents.

    • GenAI
    • LLMs
    • Prompting
    • NLP

Writing

Notes on data, lending and AI.

All writing

LLMs and AI for your daily use

Prompting better, mitigating the models refined and trained on averages, future and adaptation

Read the article

An ELO-rating model for loan default prediction

Scoring loan applicants the way chess rates players — against a model applicant who repaid — with a confidence interval that says when to buy a credit report.

Read the article

Signal over Noise

A build. A number. A good read.

Occasional notes on FinTech, data and plain writing: one thing I built, one number worth knowing with its source, and one thing worth reading. Sent when there is something worth sending.

Free and occasional. Email confirmation first, one click to unsubscribe, and no tracking pixels. Privacy

Contact

A useful conversation starts here.

Lending software, data work, AI, or something you read here. Leave your email and I’ll reply to you personally.

Your name, email and message are kept so I can reply, and nothing else is done with them. Privacy