When you advise on AI in financial services, you learn to talk fluently about risk, compliance, and explainability. When you co-found a fintech and put your own balance sheet behind the predictions, you learn what those words actually cost.
Trap one: assuming clean data exists somewhere. It does not. Every invoice we ingested arrived in a different format, with missing fields, inconsistent currencies, and counterparty names that drifted across spellings. We spent more engineering hours on canonicalisation than on the model itself, and that was the right ratio.
Trap two: optimising for accuracy instead of calibration. An 85% accurate model that is overconfident on the 15% it gets wrong will lose more money than a 75% accurate model that knows when to abstain. We rebuilt our scoring system around calibrated probabilities and rejection thresholds, not raw AUC.
Trap three: treating explainability as a UX layer. Underwriters do not need a SHAP plot. They need a sentence that survives a conversation with a client whose application was declined. We learned to generate those sentences, validate them with credit officers, and treat them as a first-class output of the model.
Trap four: underestimating the regulatory clock. Every jurisdiction we entered added six to nine months of work that no investor deck ever budgets for. Build the compliance scaffolding before you need it; retrofitting it under pressure is twice as expensive and half as defensible.
Trap five: confusing pilots with product. A successful pilot proves the model works on yesterday’s data. A successful product survives a counterparty default, a regulatory update, and a key engineer leaving all in the same quarter. We rebuilt RiskFin-IQ twice before it was ready to be called a product.
Co-founding is not consulting with skin in the game. It is a different discipline. The lessons travel back into every advisory engagement we now run.
