---
canonical: https://tanveerriaz.me/blog/banking-shipping-ai
format: text/markdown
---

People ask how a banking career and a pile of AI side projects fit together. Fair question. From the outside they look like two different worlds - regulated enterprise by day, local Ollama experiments by night. They're not. The day job taught me how to ship. The night work is where I apply it at full speed.

## Regulated environments train instincts, not just process

Twenty years at a global bank across Asia and the Middle East means you learn what "done" actually means in a regulated context. Not demo done. Not prototype done. Audit-trail done. When I build MailMind - email intelligence that never leaves your machine - that privacy-by-architecture choice isn't trendy. It's how you think when you've spent two decades around sensitive data.

Same with Sprint Calc and SalesFlow. They're side projects, but they're built with banking delivery instincts: clear inputs, visible outputs, no magic boxes. If a sprint calculator can't explain its capacity math, a delivery lead won't trust it. AI products need the same transparency.

## Sprint discipline beats model hype

In enterprise delivery, the model - whatever "model" means in your context - is never the whole story. Scope, dependencies, capacity, and who signs off matter more than the cleverest architecture slide. AI building has the same shape. Everyone asks "which model is best?" when they should ask "what job is this model doing, and what's the harness around it?"

That's why I ship Skillz as procedures, not prompts. An agent skill is a sprint board for AI work: triggers, checklists, safety rules, exit criteria. It turns "the model said something smart" into "the workflow completed reliably." Banking teams understand that. So do serious builders.

## Ship first, polish later - but know what "ship" means

My site mantra is ship first, polish later. In banking that doesn't mean reckless. It means get something working in front of real users fast, then iterate with evidence. HalalEats SG started as a chat-first halal discovery experiment, not a six-month roadmap deck. The restaurant pivot in July 2026 came from real usage, not a strategy offsite.

Enterprise peers sometimes hear "move fast" and think "move careless." The distinction: ship a thin vertical slice that proves the riskiest assumption, measure it, then expand. That's how good fintech delivery works. That's how good AI product work works too.

## Why FinTech peers should care about local-first AI

Cloud AI subscriptions are fine for exploration. They're a hard sell for regulated workflows at scale - cost, data residency, vendor lock-in, and "where did my document go?" anxiety. Local models handle 40–50% of daily knowledge work at zero marginal cost on hardware you already own. I demo this in community sessions because the finance and property crowd gets the privacy angle immediately.

The enterprise opportunity isn't "replace your bank with Ollama." It's hybrid: frontier models for the hard 50%, local models and agent harnesses for the repeatable 50%, with clear boundaries. That's a architecture conversation FinTech leaders can actually have with their teams.

## How to engage

Happy to compare notes in a personal capacity - side projects and free community sessions only, not paid consulting or work on behalf of my employer. [Connect on LinkedIn](https://www.linkedin.com/in/tanveerriaz) if you're a FinTech peer exploring delivery tooling or agent architecture. If you're a builder comparing notes on agents and local-first systems, start with [the Local AI session notes](/blog/local-ai) or browse [Skillz on GitHub](https://github.com/tanveerriaz/Skillz).

Two decades in banking didn't slow down the builder. It gave the builder better judgment about what actually ships.

[Also on LinkedIn →](https://www.linkedin.com/in/tanveerriaz)