Kaveen Hyacinth · full-stack engineer · AI-native products · Colombo
I'm a full-stack engineer. I take the AI part of a product, like the chat, the search or the assistant, and build it properly: the UI, the API, the prompts, the evals, on whichever TypeScript backend fits (Next.js, Express or NestJS), over REST or GraphQL.
Five years, four teams, four countries. Fintech screens that couldn’t be wrong, visual builders that had to feel simple, platforms that had to scale with the team. That’s the part of the job AI doesn’t replace. It’s the part AI depends on.
Foundation year at NSBM, then a First Class degree. Went straight from graduation into shipping.
Learned that money UIs have zero tolerance for "almost right". React and React Native, from associate to engineer.
Led Structure Builder 2.0 and Logic Builder, with drag-and-drop flows and logic blocks. Turns out designing workflows is great practice for designing agent pipelines.
NestJS and AWS backend services part-time; full-time, led the frontend team migrating a large platform to server-rendered Vue and Nuxt.
Architected a React, Next.js and TypeScript B2B operations platform: 20+ modules for field workflows, on a frontend built to scale with data-heavy tooling.
Users started expecting to ask, not click. Teams wanted assistants, summaries, search that understands. Calling a model turned out to be the easy part. It’s an API. The hard part is everything around it, and that’s ordinary full-stack work done carefully.
I’d rather tell you here than in the interview.
“The model is an API call. The product is everything around it.”
Six steps, any TypeScript backend. Scroll, and the diagram fills in as we go.
// app/assistant/page.tsx
const { messages, sendMessage } = useChat();What does the user actually need to do faster? Most of the work is deciding where AI belongs, and where a plain button is better. Then a streaming UI with real loading, error and empty states.
// app/assistant/page.tsx
const { messages, sendMessage } = useChat();The Vercel AI SDK gives one interface over Anthropic and OpenAI, so switching models is a config change, not a rewrite. Streaming, retries and timeouts live in a normal API handler: a Next.js route, an Express endpoint or a NestJS controller. Stream it over SSE, or through a GraphQL subscription if that’s your API style.
// assistant.controller.ts
const result = streamText({
model: anthropic(MODEL),
messages,
});
return result.toUIMessageStreamResponse();Prompts get versions, reviews and a label for what’s live. I iterate on them against saved real inputs instead of vibes, and roll back when a change makes things worse.
// lib/prompts.ts
const prompt = await langfuse.getPrompt('triage', undefined,
{ label: 'production' });RAG over the product’s own data: chunk, embed, store in Postgres with pgvector (or a document store like MongoDB), retrieve the few things that matter. Answers carry the ids of what they used, so the UI can show sources.
// lib/retrieve.ts
const docs = await retrieve(query, { k: 5 });
// pgvector · cosine · return ids to citeAgent pipelines are typed functions the model can call, a hard cap on steps, and a human confirmation before anything writes. Orchestration is mostly careful plumbing.
// assistant.service.ts
tools: { searchJobs, createTicket },
stopWhen: stepCountIs(5),Every request is traced. A dataset of real cases runs as evals before a prompt or model change ships, so “it feels better” becomes a number. Then it ships behind a flag.
// instrumentation.ts
experimental_telemetry: { isEnabled: true },
// → traces, datasets, evals in Langfuse or an alternative like Braintrust/LangSmithBuilt solo, spec first: a PRD with acceptance criteria, then AI coding agents working against it. Two are in daily use; the third is the chatbot in the corner of this page.
Short notes on frontend architecture now; AI integration notes as I learn them. Longer essays go to the newsletter.
I'm looking for a full-stack role on a team building AI-native products, somewhere I can own a feature from the database schema to the streaming UI, and make it trustworthy.
or press ~ and type contact