Riajul Islam
← All writing
AI · 16 posts

AI-accelerated development

How I actually build with AI — what gets generated, what gets thrown away, and the review discipline in between.

AI11 min read

Prompt Versioning and Why Prompts Belong in Git

A prompt is code. Storing it in a dashboard means no review, no history, and no way to tell what changed when quality dropped.

AI11 min read

Semantic Search That Understands What People Meant

Why keyword search fails on natural questions, what semantic search costs to add to an existing product, and when it is the wrong answer.

AI12 min read

Evaluating an AI Feature Before You Trust It

A test set, a scoring method and a regression check — so a prompt change cannot quietly make your answers worse without anyone noticing.

AI11 min read

Rate Limiting an AI Feature So One User Cannot Drain the Budget

Per-user, per-org and global limits for AI endpoints — and why degrading gracefully beats returning an error the user cannot act on.

AI12 min read

Document Extraction Pipelines That Survive Messy Input

Turning PDFs, scans and inconsistent forms into structured data — with validation, confidence scores and a human review path that makes it trustworthy.

AI12 min read

Provider-Agnostic AI: Swapping Models Without a Rewrite

The adapter layer that let a production app move between Claude, GPT and Groq by changing one file — and the four things it has to normalise.

AI12 min read

Semantic Search With pgvector Before You Buy a Vector Database

Postgres handles semantic search for most products. When pgvector is enough, how to index it, and where a dedicated vector database earns its cost.

AI11 min read

Streaming AI Responses Without Thrashing Your UI

Token streaming that stays smooth: batching updates, avoiding layout shift, and handling a dropped connection mid-answer.

AI11 min read

Per-User AI Cost Logging: Finding the Account Burning Your Budget

What to log on every model call, and the query that finds the one user costing more than their whole plan is worth.

AI12 min read

AI Booking Systems for Clinics, Salons and Trades

A booking assistant that checks real availability and books the slot — and the architecture rule that stops it double-booking your calendar.

AI11 min read

Validating Every LLM Tool Call With Zod

The model will eventually return a malformed tool call. The Zod schema layer that turns that from a corrupted database into a logged retry.

AI11 min read

Why AI Is a Module on a Real App, Never a Stack of Its Own

AI-first architectures fall over the moment you need auth, billing or an audit trail. The case for bolting AI onto a boring, well-built application.

AI12 min read

How Much Does It Cost to Build an AI Feature in 2026?

What an AI chat, search or extraction feature actually costs to build and to run — build fees, token spend, and the infrastructure most quotes leave out.

AI12 min read

Shipping an MVP in 19 Days Without Shipping Garbage

A real 19-day build of an AI research assistant: what got scoped out, what got generated, and the four things that never get cut no matter the deadline.

AI12 min read

How I Actually Build With AI — and Where I Still Do the Work

The honest version of an AI-accelerated workflow: what gets generated, what gets rewritten, and the review discipline that keeps fast from meaning fragile.

AI13 min read

RAG That Actually Answers: Chunking, Reranking and Citations

Most bad AI answers are bad retrieval, not a bad model. The chunking, reranking and citation changes that moved accuracy most on a production assistant.

Got a project worth
writing about?