The Databricks Forward Deployed Engineer interview loop, round by round
Evals for compound AI systems are the reported weak spot.
Often titled AI Engineer or a customer-facing engineering variant rather than FDE. Data engineering and applied GenAI on their own platform, with operational trade-offs treated as a first-class round.
How we know this. Single first-hand accounts for most rounds, and aggregator-sourced for the two data questions, which are marked as lower confidence below.
Who you meet
- Engineers
- A field or solutions lead
The rounds
1. Technical screen
What it grades
- Comfortable Python
- Data reshaping
- Pipeline thinking
2. Data and SQL
What it grades
- Monitoring retrieval failure in a RAG system
- Tokenising agent JSON logs
- Window functions, cohorts, rolling logic
The two RAG and log questions come from aggregators rather than a first-hand account, so treat them as a hint about the theme rather than as the questions you will get.
3. Applied AI and operations gating
What it grades
- How you measure whether a compound AI system works, including partial credit across multi-turn agent interactions
- Autoscaling, cold-start latency, serverless versus provisioned throughput
- Fine-tuning versus advanced prompting, argued as a trade-off
Evaluation of multi-turn agents is reported as the most common weak spot in this loop.
4. Customer case
What it grades
- Translating an architecture into business value for an executive
- Navigating customer pushback
Questions reported from this loop
- Define and defend how you would measure whether a compound AI system is working, including partial credit for multi-turn agent interactions.
- Operational trade-offs on model serving: autoscaling policy, cold-start latency, serverless versus provisioned throughput on cost.
- Translate a complex architecture into business value for a non-technical executive. Then give an example of navigating customer pushback.
- Spark, SQL, data modelling, MLflow, lakehouse architecture, and RAG over enterprise data. A customer-side workshop and notebook collaboration is part of the loop.
- Monitor retrieval failure in a RAG system. Tokenise agent JSON logs.
What is not on this page: the practice mapped to each round, and the list of what is not worth preparing for this loop. Both are in Rung.
Open this loop in Rung