AI Integration Engineer

About the role

What you'll do

• Design and ship LLM-backed features in production: prompt architecture, tool/function calling, retrieval, and structured output.
• Build the evaluation pipeline that decides whether a prompt change is an improvement — golden sets, regression runs, and human review where automation cannot reach.
• Own latency and cost budgets per feature, including caching, batching, and model-tier routing.
• Design guardrails: input validation, output constraints, refusal handling, and a deterministic fallback for when the model is unavailable.
• Handle the privacy side properly — decide what never leaves the user's environment and enforce it in the architecture, not in a policy document.
• Write up what worked and what did not, so the rest of the team can build on it.

What we're looking for

Everything in this list is a genuine requirement — if one of them is missing, the role will not work out, and we would rather say so here than at the fourth stage. We do not ask for a degree, and we do not count years of a framework that has not existed that long.

• 3+ years of software engineering, with at least one LLM-backed feature you took all the way to production.
• Strong Python or TypeScript, and comfort designing APIs around a non-deterministic dependency.
• Practical grasp of prompt engineering, embeddings, and retrieval — including where each one stops being the right tool.
• You have measured model quality with something more rigorous than reading a few outputs and deciding they looked fine.
• Working knowledge of token accounting, context limits, and how they translate into a monthly bill.
• Clear written English and the judgement to push back on an AI feature that should not exist.
• Able to work legally in Bangladesh and overlap with the team during core hours.

Nice to have

Genuinely optional. Missing all of these has never stopped us making an offer.

• Experience with local or self-hosted inference (llama.cpp, Ollama, vLLM) for privacy-constrained deployments.
• Familiarity with the Claude API, the Anthropic SDK, or MCP-based tool integrations.
• Background in classical ML or information retrieval.
• You have run a red-team pass on your own feature and found something.

What you'll work with

• Python
• TypeScript
• Claude API
• OpenAI API
• PostgreSQL + pgvector
• Node.js
• Docker

Working arrangement

Compensation

How hiring works for this role

Two to three weeks, start to offer. The practical exercise is paid, and you get an answer at every stage — including if the answer is no.

1. 1

   Application review3–5 working days

   A human reads every application — no keyword filter sits in front of us. You get an answer either way, including if the answer is no.

2. 2

   Intro call30 minutes

   A conversation with the hiring manager about your work, what you want next, and the specifics of the role — including compensation, so nobody's time is wasted later.

3. 3

   Practical exercise4–6 hours, paid

   A scoped, realistic task close to the actual work — or, for design and motion roles, a walkthrough of work you already own. We pay for take-home time. We do not ask for free work on a live problem.

4. 4

   Team conversation45–60 minutes

   You meet the people you would work with, and the founder. We go through your exercise together, and you get as much time to ask questions as we do.

5. 5

   Decision and offerWithin 3 working days

   A written offer with salary, level, start date, and the probation terms spelled out. If it is a no, you get specific feedback rather than a template.

How to apply

Equal opportunity

Degird hires on the strength of your work. We do not screen on gender, religion, age, ethnicity, disability, marital status, or where you went to school — and we will make reasonable adjustments to any stage of our process if you tell us what would help. If something in this posting is a barrier rather than a requirement, say so in your application and we will look at it.