Joe Kariuki

founder & engineer @devbrewai, building @calidorai

[Joe Kariuki @joekariuki]

Buying a home is the biggest purchase most people ever make, and the deal moves by hand in WhatsApp and email, whatever CRM sits behind it.

I'm building something better, and I'll share more soon.

If you know a real estate developer, I'd love an intro.

[Joe Kariuki @joekariuki]

Repo link:
https://t.co/Nriz3Pbevc

[Joe Kariuki @joekariuki]

Sentinel is live.

An open source case study by Devbrew (@devbrewai) demonstrating real-time AI fraud detection and sanctions screening for cross-border payments, behind a single API.

Live demo and repo links in replies.

[Joe Kariuki @joekariuki]

Live demo:
https://t.co/qcZlt3hjiE

[Joe Kariuki @joekariuki]

You can't automate trust, but you can teach it to learn.

[Joe Kariuki @joekariuki]

Every payment company becomes its data pipeline. Eventually.

[Joe Kariuki @joekariuki]

The market doesn't reward truth. It rewards conviction.

[Joe Kariuki @joekariuki]

The less human the payment process becomes, the more human the experience will feel.

[Joe Kariuki @joekariuki]

Complexity scales faster than intelligence.
Fix the complexity first.

[Joe Kariuki @joekariuki]

Most AI projects fail because they automate noise, not signal.

[Joe Kariuki @joekariuki]

"AI-powered" means nothing to investors.

“AI reduced our cost per transaction by 35%" gets you funded.

Show the math, not the buzzword.

[Joe Kariuki @joekariuki]

Reconciliation is basically pattern matching plus exception handling at scale.

ML can learn match likelihood, and an agent can gather supporting context and draft the exception notes.

Your team keeps the final call, but the busywork drops hard.

[Joe Kariuki @joekariuki]

A tiny team with strong tooling can do absurd output.

The constraint is rarely talent. It’s convincing yourself that being “busy” means you’re being productive.

Pick one painful workflow, instrument it end to end, automate the boring parts, then ship.

[Joe Kariuki @joekariuki]

The model is rarely the bottleneck. The workflow is.

AI is most useful when it sits inside the decision loop: recommend, explain, log rationale, trigger next steps.

That is how you get speed without losing control or auditability.

[Joe Kariuki @joekariuki]

A lot of payments pain is data quality, not modeling.

AI can help by monitoring streams for anomalies, drift, and broken partner feeds, then summarizing what changed and where to look.

It turns “why are numbers off” into a fast, searchable answer.

[Joe Kariuki @joekariuki]

Fraud detection gets a lot better when you stop treating events as isolated.

Graph based features plus ML can flag suspicious networks, not just single transactions.

That means earlier detection, fewer chargebacks, and less investigator time wasted on noise.

[Joe Kariuki @joekariuki]

Open source research case study, building in public.

We’re in the final stretch.

What dashboard feature would your risk or compliance team use day one? Reply with it.

Repo: https://t.co/UYh17ybnsW

[Joe Kariuki @joekariuki]

Built a fraud + sanctions scoring API for cross border payments at @devbrewai

End-to-end: ~32ms clean, ~48ms fraud (target <200ms p95).

At 1M txns/mo, shifting 1% out of manual review = 10k fewer reviews, roughly $50k/mo at $5 per review.

If you run payments risk/compliance and want this in your stack, DM me.

[Joe Kariuki @joekariuki]

If you manage multi currency settlement, static thresholds are usually the hidden reason you keep too much idle float.

A simple ML forecast of inflows and outflows lets you pre fund with more confidence.

The result: less idle capital, fewer surprises, calmer ops.

[Joe Kariuki @joekariuki]

Most teams “scale compliance” by hiring.

A more scalable path is a risk based flow: low risk auto clear, gray area gets prioritized, obvious risk gets blocked.

ML helps by scoring risk and packaging evidence for reviewers, so your team focuses on spending time where it matters.