AI Graduate Fall Recruitment Review: 8 Companies Interviewed, 3 Offers Received

Background

Let me introduce myself: 985 university undergrad and master's, mathematics undergrad, computer vision master's focusing on object detection research. Fall recruitment ran from September submissions to December signing — interviewed at 8 companies, ultimately received 3 offers. The whole process had ups and downs, surprises and regrets. Let me do a complete review today.

My goal was clear: AI algorithm engineer, leaning toward CV, but open to NLP and recommendation directions. Companies I applied to covered Big Tech AI Labs, AI unicorns, and autonomous driving companies. Let me walk through the timeline in detail.

Interview Process Review

Preparation Strategy

I started preparing in July, spending about two months. Preparation had three phases:

Phase 1 (July): Knowledge review + coding practice. Focused on ML fundamentals (SVM, Random Forest, XGBoost), DL fundamentals (CNN, RNN, Transformer, optimizers), and CV knowledge (object detection, image segmentation, GANs). Coding: 3-5 LeetCode problems daily, focusing on Hot 100.

Phase 2 (August): Project review + paper preparation. Went through my research projects from start to finish, prepared 3 versions of paper presentations (1-minute, 5-minute, 15-minute). Also started reading interview experiences to understand each company's style.

Phase 3 (Early September): Mock interviews + filling gaps. Did 3 mock interviews with seniors, discovering expression issues — like being too verbose when presenting papers and not being structured enough in answers.

ByteDance AI Lab (Failed Round 2)

ByteDance was my first interview, mid-September. Round 1 went smoothly — all fundamentals: ResNet structure and improvements, Faster R-CNN pipeline, YOLO vs SSD differences, Transformer Self-Attention computation. Coding: implement NMS (Non-Maximum Suppression) from scratch, which I wrote quickly. Good interviewer feedback.

Round 2 was a disaster. The interviewer asked an open-ended question: How would you design a real-time object detection system running on edge devices? I lacked a system design framework — jumping from point to point without clearly explaining model compression (quantization, pruning, distillation) selection logic or deployment pipeline. The interviewer followed up on TensorRT optimization principles — I only knew concepts, couldn't explain details. Finally, a behavioral question: "What's the biggest technical challenge you've faced?" My answer was too vague, not using the STAR method. Failed Round 2.

Alibaba DAMO Academy (Offer)

Alibaba was late September. Longer process but good experience. Round 1: fundamentals + project, focusing on Faster R-CNN's RPN principles, FPN's role, and data augmentation strategies in my research. Coding: implement IoU calculation from scratch — straightforward.

Round 2 was a cross-functional interview, going deeper. Asked about Vision Transformer understanding, ViT vs CNN differences, Swin Transformer's window attention mechanism. An interesting question: If your model's performance degrades online, how do you troubleshoot? I mentioned data drift, feature distribution changes, and annotation quality degradation — the interviewer was satisfied.

Round 3 was with the manager, discussing career plans and team direction — more relaxed. Finally, HR discussed salary expectations. The whole process took about 3 weeks, ultimately receiving an offer at P6 level.

Tencent AI Lab (Failed Round 3)

Tencent had the longest interview process I experienced — over a month. Passed Rounds 1 and 2: Round 1 on CV fundamentals, Round 2 on projects + papers. Round 3 was with a director, asking many open-ended questions.

Round 3 failure reason: What do you think is the most important direction in CV for the next 3 years? I said multimodal and 3D vision, but didn't go deep enough — didn't analyze from both technology trends and commercial application perspectives. Another question: If your research direction doesn't align with the team's direction, how do you handle it? My answer was too direct — "I'll adjust my direction" — without demonstrating independent thinking. Failed Round 3, quite a pity.

Baidu AI (Offer)

Baidu was in October, overall experience was average. Round 1 on fundamentals, Round 2 on projects + system design, Round 3 with technical director. Baidu emphasizes engineering ability — Round 2 had a system design question: Design an image moderation system that processes millions of images in real-time. I covered distributed architecture, model serving, message queues, and degradation strategies — the interviewer approved.

Baidu's interview style is pragmatic, not chasing cutting-edge topics, valuing engineering implementation ability more. Received an offer at T5 level.

SenseTime (Offer)

SenseTime was one of my top choices — very strong in CV. Round 1 covered many CV frontiers: DETR principles and improvements, SAM (Segment Anything) architecture, Diffusion Models in image generation. Coding: implement a simple convolution operation from scratch. Round 2 deep-dived into papers, having me detail my research work and questioning experimental design rationality. Round 3 was HR, discussing salary and team.

SenseTime's interview style is very academic — interviewers are researchers, asking deep questions. Received an offer with competitive CV-direction salary.

Megvii (Failed Round 2)

Megvii Round 2 failure was due to a coding problem I couldn't solve. The interviewer asked me to implement a Transformer Encoder Layer from scratch, including Self-Attention + FFN + Residual + Layer Norm. I mixed up matrix dimensions in Self-Attention and couldn't debug it in time. Lesson learned: implementing ML algorithms must be practiced to muscle memory.

NIO Autonomous Driving (Withdrew)

NIO offered an interview opportunity, but I learned their autonomous driving team leans engineering-heavy with limited algorithm innovation space, and the office was in Shanghai's suburbs with inconvenient commuting. After consideration, I withdrew.

Pony.ai (Failed Round 1)

Pony.ai's Round 1 asked many 3D vision and point cloud processing questions — areas where I was underprepared. Topics like BEV perception principles, PointNet architecture, and multi-sensor fusion strategies. Failed Round 1, showing the importance of understanding target companies' technical directions before interviewing.

Interview Rhythm and Mindset Management

Fall recruitment's biggest challenge isn't technical — it's mindset. When I failed two companies in September (ByteDance, Megvii), I was really anxious, even doubting whether I was cut out for AI. Later I adjusted several things:

First, don't put all eggs in one basket. I applied to 8 companies simultaneously, so even failing a few still left opportunities.

Second, review after each failure. After ByteDance's Round 2, I spent two days specifically on system design; after Megvii, I spent a week implementing 10 ML algorithms from scratch.

Third, maintain rhythm. Fixed daily time for coding, knowledge review, and project prep — don't let failures disrupt your schedule.

Fourth, talk to people. Mutual encouragement with classmates during fall recruitment is important — don't carry it alone.

Key Questions Summary

1. ResNet structure and improvements? Why do residual connections work?

2. Faster R-CNN's RPN principles? Anchor design?

3. YOLO vs SSD differences? Improvements across YOLO versions?

4. Transformer Self-Attention computation flow?

5. Implement NMS (Non-Maximum Suppression) from scratch

6. Design a real-time object detection system (edge device deployment)

7. Vision Transformer vs CNN differences?

8. Swin Transformer's window attention mechanism?

9. How to troubleshoot online model performance degradation?

10. Most important CV direction for the next 3 years?

11. Implement IoU calculation from scratch

12. Design an image moderation system (million-level real-time processing)

13. DETR principles and improvements?

14. SAM (Segment Anything) architecture?

15. Implement Transformer Encoder Layer from scratch

16. BEV perception principles? PointNet architecture?

Insights and Advice

1. System design preparation is essential: AI roles no longer just ask algorithm principles — system design questions are increasing. Learn a framework (from requirements analysis to architecture design to trade-off analysis), then practice 3-5 typical problems.

2. Implementing ML algorithms from scratch is hard skill: NMS, IoU, Self-Attention, convolution operations, K-Means — you must be able to implement them. Not just understanding, but writing and running them within 30 minutes.

3. Paper presentation needs structure: Don't present chronologically — use "Problem-Method-Innovation-Experiments-Limitations" structure. Prepare 1-minute, 5-minute, and 15-minute versions.

4. Understand target companies' technical directions: Prepare 3D vision for autonomous driving companies, ranking models for recommendation companies, frontier papers for CV companies. Blind preparation wastes effort.

5. Use STAR method for behavioral interviews: Situation-Task-Action-Result — structure every story this way for clarity and persuasiveness.

6. Mindset determines success: Fall recruitment is a marathon, not a sprint. Maintain rhythm, review promptly, and don't self-doubt.

FAQ

Q: What degree is needed for AI campus recruitment?
A: Big Tech AI Labs generally require a master's minimum; undergrads can get in but face stiffer competition. PhDs have advantages but aren't required.

Q: Can you join Big Tech AI Labs without top conference papers?
A: Yes, but you need highlights in other areas — like Kaggle gold medals, open-source projects, or internship experience. Papers are a plus, not a requirement.

Q: When should I start preparing for fall recruitment?
A: Start in July, complete basic prep by August, submit in September. Early batches start June-July — follow target companies' recruitment updates.

Q: Which is better for employment — CV or NLP?
A: In 2026, NLP has more positions due to LLMs. But CV demand is also strong in autonomous driving and industrial inspection. Choose based on personal interest.

Q: Can I reapply after failing an interview?
A: Most companies have a cooling period (3-6 months). After that, you can reapply. If you fail fall recruitment, you can try again in spring.