Interview Prep — Behavioral Questions

STAR format: Situation → Task → Action → Result Prepare 2-3 specific stories for each theme. Numbers and outcomes matter.

Theme 1: Handling Model Failures in Production

Sample Question

"Tell me about a time a model you deployed caused a problem in production."

STAR Template

Situation: Describe the model, deployment context, and what went wrong.
Task: What was your responsibility when the issue occurred?
Action: Walk through your debugging process step by step.
Result: Quantify the impact and what you changed to prevent recurrence.

Strong Answer Framework

1. State the failure mode clearly: drift, edge case, wrong metric
2. Describe your monitoring that caught it (or didn't)
3. Explain your root cause analysis
4. Describe the fix: model update, fallback rule, threshold change
5. Describe what you put in place afterward

Example Talking Points

• "We deployed a person detection model for access control. After 3 weeks, the FPR spiked from 0.2% to 4% — we investigated and found the camera's autofocus behavior had changed after a firmware update, introducing motion blur we hadn't seen in training. We added online evaluation against a held-out labeled set that ran every 6 hours, and added blur detection to the preprocessing pipeline to reject low-quality frames."
• "Our object detection AP dropped from 0.72 to 0.58 silently over a month. We had no input distribution monitoring. After this, I implemented data drift detection using KL divergence on predicted class distributions compared to training distribution."

Key Points to Hit

• Show ownership: you caught it or helped catch it
• Show engineering rigor: systematic debugging, not guessing
• Show learning: what monitoring/process you added afterward

Theme 2: Technical Leadership & Cross-Team Collaboration

Sample Questions

• "Describe a time you influenced a team that wasn't directly reporting to you."
• "Tell me about a time you drove a technical decision that was controversial."

STAR Template

Situation: Team structure, competing priorities, technical disagreement.
Task: What outcome did you need to achieve?
Action: How did you make your case? What data did you use?
Result: Was your approach adopted? What was the business impact?

Strong Answer Framework

1. Acknowledge the competing viewpoint fairly
2. Describe the analysis or prototype you built to make your case
3. Describe how you communicated it (design doc, A/B test, demo)
4. Note how you handled disagreement professionally

Example Talking Points

• "We had a debate about whether to use YOLOv8 or a two-stage detector for our warehouse robots. The product team wanted accuracy; infra team wanted to keep latency under 50ms. I ran a two-week spike with both models on our actual hardware, documented the Pareto frontier of accuracy vs latency, and proposed YOLOv8m with TensorRT. This data-driven approach won over both teams."
• "I proposed migrating our inference stack to Triton Inference Server. The engineering team was skeptical of the migration effort. I built a proof-of-concept over a weekend, showed 3× throughput improvement, and documented the migration path step-by-step to reduce risk perception."

Theme 3: Dealing with Ambiguous Requirements

Sample Questions

• "Tell me about a time you had to make a decision without all the information you needed."
• "Describe a project where requirements changed significantly mid-way."

STAR Template

Situation: What was unclear? What were the risks of getting it wrong?
Task: What did you need to deliver and by when?
Action: How did you structure the ambiguity? What questions did you ask?
Result: How did the project turn out? What would you do differently?

Strong Answer Framework

1. Show you actively reduced ambiguity rather than waiting
2. Describe how you defined the MVP and deferred non-essential work
3. Show how you managed stakeholder expectations around uncertainty
4. Highlight what you learned about scoping

Example Talking Points

• "We were asked to 'improve the detection accuracy' on a manufacturing line, with no baseline metric, no labeled dataset, and no definition of success. I spent the first week establishing baselines — ran our existing model, collected 500 labeled ground-truth frames from the line, and wrote a one-pager defining what 'good' meant: mAP@0.5 > 0.80 at < 50ms latency. This became the success criteria the whole team aligned on."
• "Mid-project, the product team changed the target from 5 classes to 12. I flagged that this would require 5× more labeled data and an additional 3 weeks. We agreed to release v1 with 5 classes and v2 with all 12. This prevented a slip while keeping momentum."

Theme 4: Technical Deep-Dives & Problem Solving

Sample Questions

• "Walk me through the most technically challenging project you've worked on."
• "Describe a time you had to learn a new technology quickly."

STAR Template

Situation: What was the hard technical problem?
Task: What did success look like?
Action: What was your approach to breaking down the problem?
Result: What did you achieve? What did you learn?

Strong Answer Framework

1. Be specific — don't generalize ("I worked on computer vision")
2. Explain why it was hard (technical, not just time pressure)
3. Show systematic problem-solving: hypothesis → experiment → conclusion
4. Quantify the improvement

Example Talking Points

• "I had to optimize a segmentation model to run at 30 FPS on an embedded NVIDIA Orin. Starting point was 8 FPS. I profiled with nsys and found 60% of time was spent in the decoder upsampling. I replaced bilinear + conv with a lightweight learned upsampler, exported with TensorRT FP16, and achieved 34 FPS — a 4.25× improvement."
• "I needed to understand RAFT (optical flow) for a video stabilization project in 3 days. I read the paper, ran the official code, then re-implemented the correlation volume from scratch. Understanding it from first principles let me debug a numerical precision bug that the pretrained weights obscured."

Theme 5: Mentorship & Growing Others

Sample Questions

• "Tell me about a time you mentored a junior engineer."
• "Describe how you've contributed to your team's technical growth."

STAR Template

Situation: Who were you mentoring? What was their challenge?
Task: What were you trying to help them achieve?
Action: What specific steps did you take?
Result: What progress did they make?

Example Talking Points

• "A junior engineer on my team was struggling to get a detection model to converge. Rather than debugging for them, I sat down and taught them how to read loss curves and gradient norms systematically. I showed them how to first overfit a single batch, then scale up. Within a week they were self-sufficient with training debugging."
• "I wrote a 'Model Debugging Checklist' for my team: 10 questions to answer before escalating a training problem. It reduced the time from 'model not working' to 'root cause found' by about 60%."

Questions to Ask the Interviewer

Technical Questions

• "What does the model deployment pipeline look like today? What's the main bottleneck?"
• "How do you evaluate model drift in production? What triggers a re-train?"
• "What's the typical ratio of data labeling / model training / deployment work for the team?"
• "What's the hardest CV problem you're currently trying to solve?"

Team & Culture

• "How does the team handle disagreements on technical direction?"
• "What does career growth look like for a CV/ML engineer here?"
• "What are the biggest technical challenges the team will face in the next 12 months?"

Process

• "How long does a new model typically take from first experiment to production?"
• "How do you balance shipping quickly vs building maintainable systems?"

Negotiation & Offer Notes

• Always negotiate. The first offer is rarely the final offer.
• Anchors: competing offers, market data (levels.fyi, blind), your current package
• For ML engineer roles, equity + bonus often exceed base for senior levels — ask for details
• Negotiate title and scope too — "Senior ML Engineer" vs "Staff" is a big career difference
• Get promises in writing (team, project, compute budget)