About the role

You will build products on top of language models the way production software is built: with an understanding of how the model actually behaves, evals instead of vibes, and the engineering discipline of a specialist in at least one niche.

AI coding has blurred the line between niche specialists and generalists. What still separates people who can ship from people who cannot is a strong base in one domain plus real software-engineering principles — with today’s tools, that combination does wonders. That is who this posting is written for. One posting covers two levels, Engineer and Senior; the level is decided in the process, not by the years on a CV.

What you will do

Design and ship LLM-backed systems end to end — agents, workflows, tool use, retrieval — on Cloudflare Workers, TypeScript and Terraform, the stack every Scaledex system runs on.
Own the harness: Claude Code skills, agents, hooks and memory; tool design; context and cost budgets; how a system recovers when the model is wrong.
Build evals and gates before features: golden sets, adversarial checks, regression runs on every prompt change.
Read model behaviour closely — where it hallucinates, when it over-complies, what it does under long context — and design around it rather than hoping.
Review code, write design notes, and explain what you built on the blog in the house voice.

What we look for

Required

Software you shipped and kept running in one niche — web platforms, data systems, mobile, infrastructure — with the principles that come from that: tests, types, observability, boundaries.
Hands-on LLM work you can show: a system with tools or agents, how you evaluated it, what broke in production and what you changed.
Fluency with an agentic coding harness (Claude Code preferred): skills, agents, hooks, memory, and an opinion on where they stop helping.
TypeScript in earnest; Cloudflare Workers, D1, R2 or an equivalent serverless platform; Git as a daily tool.

Useful

Terraform or another infrastructure-as-code tool; Python for evals and data work; a public write-up or talk; Bangla for client conversations.

Not for

People whose only path to working code is prompting until it runs. You will be asked to debug, to test non-deterministic outputs, and to say when a model should not be used at all.

How we assess

Apply with the form: the system you built, a repository you are proud of, and a few short answers. About 30 minutes.
A reply within five working days, yes or no, with a reason.
One conversation of 45 minutes about that system and one debugging story.
A short exercise — half a day, paid — on a real harness from our repository, or a two-hour live pairing session; your choice.
An offer within a week of the exercise, after references.

What we offer

A monthly salary agreed at offer for the level you come in at; tell us your expectation in the form and we will tell you ours in the first conversation.
Hybrid work with a desk in Gulshan-1, the equipment you need, a budgeted allowance for AI tokens, and a learning budget.
Ownership from the first conversation to the running system, in a company that writes about what it builds and teaches it in its workshops.

How to apply

Use the application form. It saves as you go, and you get a reference and a confirmation by email the moment it is in.