Journal
What we learn, written down.
Engineering notes from building AI apps in production, mostly from Finly and Investolio, plus what we are reading along the way.
Field notes
From the code we run.
What we have learned building and running Finly and Investolio.
- Receipt scanning in productionOne vision-model call reads the receipt. A surprising amount of code around it makes the result something you can trust. How Finly's scanner works, and where it broke.
- Confirm before you write: a tool-calling assistant for moneyTelling a model to ask first is not enforcement. How Finly's chat assistant is built so that nothing is written until you tap confirm.
- A market-data service with no API keysInvestolio gets its prices, dividends and exchange rates from free, keyless sources, with a small service of ours in front and a fallback behind every path.
- Shipping an Expo app with self-hosted OTA updatesJavaScript changes to Finly reach phones without a store review, through an update server we host. How the setup works, and the rules that keep it safe.
Reading notes
What other people found.
Our summaries of research, industry reports and customer stories about AI at work, with sources linked.
- Where Jev fits in business automation.Essay
- Decision models vs chat models in a real workflow.Essay
- What Jev is — and why it spread so fast.Research
- Two years of AI ROI: what moved.Research
- Learn It Live cut tickets 40%.Industry brief
- Writing tasks got faster—and better.Research
- Sprint-based AI delivery.Essay
- Agent demos hide the hard parts.Essay
- Local models when data cannot leave.Essay
- The cost of retries and review.Essay
- 88% use AI. Few have scaled it.Research
- Shadow AI needs a lighter governance model.Essay
- Only 5% get AI value at scale.Research
- RAG fails when retrieval is wrong.Essay
- HackerOne’s scheduling automation ROI.Industry brief
- Keep a person in the loop.Essay
- Rebrandly cut support tickets in half.Industry brief
- AI Index 2025: what businesses should notice.Research
- When not to use AI.Essay
- Inference costs fell more than 280×.Research
- Litmus consolidated support. Retention rose.Industry brief
- One workflow. Then a platform.Essay
- Gen AI use doubled. EBIT impact stayed rare.Research
- The prototype is not the product.Essay
- Lab speed is not field pull requests.Research
- Support assistants raised resolution 14%.Research
- Measure the work. Not the tokens.Essay
- Most AI pilots stall before value.Research