Track 10 · sessions 234–239
Capstone and interviews
Take apart a modern model, build a sports key-moment search system, practice interview design questions.
Capstone
Read a modern model end to end, then design, build and evaluate a sports key-moment search system.
234Ready
Dissect a modern open model
- Pick one, for example Gemma 4, Qwen3, DeepSeek-V3 or Qwen2.5-VL
- From the config: layers, heads, KV heads, experts, vision tower, context length
- Compute parameters, FLOPs per token, KV cache per token, memory to serve
- Read its technical report and explain every design choice with what you learned
235Ready
Capstone design: sports key-moment search
- Requirements: queries, latency, accuracy, cost, hardware (L4 + TensorRT)
- Design: segmenting, embedding, indexing, VLM re-ranking, timestamps
- Latency and cost math before building
- Risks, failure modes, and what to measure
236Ready
capstone build
- Build the index and the retrieval service
- Add VLM re-ranking and timestamp refinement
- Serve with vLLM or TensorRT-LLM; profile and fix the bottleneck
237Ready
capstone evaluation and failure analysis
- Evaluate retrieval and grounding on your own labeled set
- Error analysis: group the failures by cause
- Write a short technical report with numbers, limits and next steps
238Ready
Interview drills: ML system design for LLM, VLM and video services
- A repeatable answer frame: requirements, numbers, design, bottleneck, failure modes
- Drill 1: a chat API with a latency target
- Drill 2: natural-language video search over a sports archive
- Drill 3: real-time key-moment alerts on live broadcast
239Ready
Final review and what to read next
- Review the 7 parts with one hard question each
- Open problems: long video, efficient reasoning, unified models, evaluation
- Your reading list for the next 6 months
- How to keep up: surveys, conferences, code