01 / APPLIED AI AND EVALUATION
AI Engineer
Make AI useful when the questions get difficult.
STEP INSIDE THE WORK
AN ILLUSTRATIVE PROJECTAn answer is only as useful as the evidence behind it.
A policy assistant sounds convincing. Some of its answers are wrong. Where do you begin?
YOUR MOVE / 01
Start with the failure.
Read the questions people actually ask. Find where the right information was missing, where retrieval failed, and where the answer went beyond its sources. Define a baseline before changing the model.
WHAT YOU MAKE VISIBLE
Real user questions
Separate retrieval from generation
A baseline you can reproduce
Think from first principles.
Make your reasoning visible.
01 / YOUR MISSION
The demo is the beginning. Build what earns trust after it.
An AI assistant gives a brilliant answer. Then a customer asks the question nobody tested. That is where this role gets interesting. At devx, you will turn models, data and experiments into useful AI capabilities for real workflows. Bring curiosity about how things work and the discipline to test your own ideas. Your contribution is an improvement you can explain, reproduce and defend with evidence.
The quality and behaviour of an AI component: how it retrieves evidence, produces an answer or uses an approved tool. You partner with application engineers to put that capability into a usable product.
We reimagine customer interactions, business operations and enterprise architecture around AI. You bring fresh questions. Experienced practitioners bring context and judgment. The work connects both to a real customer outcome.
02 / WHAT YOU WILL WORK ON
Make your curiosity
useful.
Build useful AI components
Write Python for scoped tasks such as extraction, classification, semantic search or assisted decision-making. Start with a simple baseline and make each experiment answer a clear question.
Connect answers to evidence
Develop retrieval-augmented generation workflows over approved information. Prepare documents, compare retrieval approaches and attach supporting sources. Respect document access rules and evaluate retrieved evidence separately from the generated answer.
Make quality measurable
Build representative evaluation cases with the team. Compare task accuracy, unsupported answers, response time and usage cost, including difficult questions and cases the system should decline.
Design for failure
Build bounded tool-using workflows where they serve the task. Validate structured outputs, restrict available actions, handle timeouts and stop conditions, and require reviewed human approval for consequential steps. Investigate unsupported answers and prompt-injection attempts with the team.
Make experiments usable by others
Work with data, application and cloud engineers to integrate your component. Contribute tests, experiment records and clear handover notes so improvements survive beyond your notebook.
What good work looks like
- An experiment another engineer can reproduce, with representative cases and a meaningful baseline.
- An improvement evaluated for task quality, failure cases, latency and usage cost.
- A component with documented limits, access controls and a tested fallback.
Use approved AI tools to explore approaches, draft experiments and investigate failures. Verify the code and evidence yourself, keep evaluation examples separate from tuning, and record what changed your conclusion.
03 / YOUR STARTING POINT
Bring a foundation.
Build the range.
Coursework, personal projects, research and student initiatives all count. Previous full-time experience is not required.
- Python you can read, write and debug, including functions, common data structures and library use.
- Introductory ML, probability and statistics: training versus evaluation, overfitting, and choosing a useful metric.
- A coursework, personal or research project involving ML, information retrieval or an LLM. Explain the data, baseline, evaluation method and limitations; a carefully evaluated small project is enough to start.
- A habit of questioning results, recording experiments and explaining both your contribution and its limitations.
Useful exposure
Git, SQL, pandas, scikit-learn or PyTorch; embeddings and model APIs. These are useful starting points, not a requirement to master every tool or train a large model.
04 / HOW YOUR OWNERSHIP GROWS
Learn in the work.
Grow through the feedback.
Start with a bounded experiment or component and experienced review. As your judgment develops, take more responsibility for its evaluation, integration and behaviour in use. Build the foundations for applied AI and ML engineering by learning when a simpler method is enough and when extra complexity is justified.
Explore
Understand the problem and ask useful questions.
Build
Make a defined contribution and learn through review.
Own
Take on broader responsibility as your readiness grows.
05 / THE DEVX CULTURE
The principles
show up in the work.
Understand what the customer is trying to change. Ask the extra question, surface the real obstacle and connect your work to their goals.
A feature shipped, but the workflow is still slow. Stay with the problem and find out why.
Client obsession. Understand what the customer is trying to change. Ask the extra question, surface the real obstacle and connect your work to their goals.
AI-native everything. Use AI to improve how you research, build, analyse and document. Understand its limits, verify the output and handle customer information responsibly.
Hire and train exceptional talent. We bring young talent and experienced practitioners together. Bring ambition, seek direct feedback and put new understanding into practice.
Document to scale. Write down the decision, its context and what you learned. Make your work understandable enough for someone else to continue, question or improve it.
Only ever be honest. Flag risk while there is time to act. Separate what you know from what you assume. Give clear feedback and make uncertainty visible.
06 / START A CONVERSATION
Show us the way
you think.
Share a résumé and one project or notebook. Explain the problem, your contribution, the baseline, how you checked the result, and one failure that changed your approach.
Join through an internship with the opportunity to convert to a full-time role. The hiring team will share internship duration, work arrangements, campus eligibility and conversion criteria during recruitment.
An illustrative project
A support team needs answers from a changing policy library. You could build one part of the assistant, create an evaluation set, and show whether a retrieval change improves supported answers. Include an outdated policy, an ambiguous question and a question with no answer. Success includes knowing when to ask for clarification or hand over.