Howsoftwareactuallyships.
Field notes on building AI into real products — RAG, chatbots, agents, and the architecture decisions that outlive a release. Written by the team that ships them.
Notes from thedelivery floor.
Retrieval pipelines that return the wrong passage with total confidence. Chatbots that pass a demo and fail a customer. The moment a model stops answering and starts acting, and what that changes about who can approve what. These are the problems that show up once an AI feature meets production, and the tradeoffs behind each one.
Alongside them: the engineering decisions that outlive a release — how to version an API before anyone depends on it, which technical debt compounds and which is safe to carry, and when blockchain is genuinely the right tool rather than a database with extra steps.
- AI in production
- RAG, retrieval quality, hallucination, agents, and prompt injection — what holds up once real users and real documents are involved.
- Architecture calls
- The decisions with a long tail: model selection, API contracts, cross-platform mobile, and the debt worth paying down.
- Industry context
- How the same technology lands differently in healthcare, fintech, EdTech, and e-commerce — where the constraints actually bind.
AI Engineering
What Is RAG (Retrieval-Augmented Generation), and When Do You Need It?
RAG lets a language model answer from your own documents instead of guessing. Here is how it works, when it beats fine-tuning, and where it quietly fails.
2 min read
EngagementHow Long Does It Take to Build a Custom Software Product?
Timelines depend on scope, not headcount. Here's what actually sets the schedule for a custom build, and why the discovery phase is where a realistic date gets decided.
1 min readAI Engineering
Building a Production Chatbot: What Separates a Demo from a Product
A chatbot demo takes an afternoon. A chatbot that survives real users takes guardrails, evaluation, escalation, and observability. Here is the gap.
1 min readAI Engineering
Vector Databases Explained: Do You Actually Need One?
Vector databases power semantic search and RAG — but plenty of projects add one before they need it. Here is what they do and when Postgres is enough.
1 min readAI Engineering
Why LLMs Hallucinate, and What Actually Reduces It
Hallucination is not a bug to be patched — it is how these models work. Here is what genuinely reduces it in a product, and what only appears to.
1 min readAI Agents vs Chatbots: What Is the Real Difference?
Everything is being rebranded as an "agent". The distinction that matters is whether the system can take actions with consequences — and what happens when it takes the wrong one.
1 min readHow to Choose an LLM for Your Product
Benchmark scores are the least useful input to this decision. Here is what actually determines which model belongs in your product.
1 min readSecurity
Prompt Injection: The Security Risk in Every LLM Feature
If your LLM feature reads untrusted text — user input, emails, web pages, uploaded documents — it can be instructed by that text. Here is what that means and how to contain it.
1 min readHealthcare
AI in Healthcare Software: Where It Helps and Where It Cannot Go
Healthcare is where AI hype meets real consequences. Here is what genuinely works in clinical software today, and the line responsible teams do not cross.
1 min readFinTech
AI in FinTech: Fraud, Onboarding, and the Cost of a False Positive
In financial software, a wrong AI decision has a price tag. Here is where machine learning genuinely pays for itself, and how to think about the error tradeoff.
1 min readEdTech
AI in EdTech: What Personalised Learning Actually Requires
"Personalised learning" is claimed by every EdTech product. Here is what it takes to build the real thing, drawn from shipping an AI-powered grammar platform.
1 min readE-Commerce
AI for E-Commerce: Search and Recommendations That Actually Convert
Most e-commerce AI spend goes to recommendations. Most e-commerce revenue is lost in search. Here is why the priority is usually backwards.
1 min readEngineering
Next.js App Router: What Actually Changed, and What It Costs You
Server Components change where your code runs, not just how it is organised. Here is the mental model, and the traps teams hit migrating.
1 min readCore Web Vitals: Which Numbers Actually Matter
Most performance work optimises the wrong metric on the wrong build. Here is what to measure, where, and what genuinely moves the score.
1 min readMobile
React Native vs Flutter: How to Actually Decide
Both ship good apps. The decision rarely comes down to the framework — it comes down to your team, your integrations, and how native your app needs to feel.
1 min readBlockchain
When Blockchain Is Actually the Right Tool
Most problems described as blockchain problems are database problems. Here is the short test for which one you have — from a team that ships both.
1 min readEngineering
API Design That Survives Its Second Year
Most API pain is not performance — it is versioning, error handling, and pagination decisions made in week one and regretted for years.
1 min readTechnical Debt: When to Pay It, and When It Is Fine
Not all technical debt is worth paying down. Here is how to tell the debt that compounds from the debt you can carry indefinitely.
1 min readDevOps
CI/CD: What a Small Team Actually Needs
You do not need a platform team to deploy safely. Here is the minimum pipeline that genuinely reduces risk, and what is premature below a certain size.
1 min readEngagement
Staff Augmentation vs Agency: Which One Do You Actually Need?
These are sold interchangeably and are not interchangeable. The difference is who owns the outcome — and that changes what you should expect.
1 min readAI Engineering
Semantic Search vs Keyword Search: What Changes for Users
Semantic search finds things keyword search misses — and misses things keyword search finds. The best systems use both. Here is why.
1 min read
Let's Start a Conversation
Reading is one thing. Tell us what you're building and we'll come back within one business day with a scoped plan.