全文共 377 字 预计阅读 2 分钟
interview

agirbt(oc 6.4)

i felt tired

2026.06.01

self-introduction

0:00 - 03:16 response: good

Q1

responsible part in energy-agent 03:17 - 17:25 response: you say a lot (but without derogatory tone)

Q2

how do u define the json schema in the mcp 17:40 - 19:12 no response

Q3

difference between mcp & function call 19:20 - 20:50 no response

Q4

does the k8s-pod experience in redis equals to meory? 20:50 - 21:35 response: okay

Q5

how to avoid the experience illusion? 21:50 - 23:56 response: en

Q6

responsible part in iot platform 24:00 - 26:45 no response

Q7

how to optimize the core interface's response for reporting? 27:05 - 30:10 ps: i forget the caffine + redis part, fk!!!! no response

Q8

how does the gateway collect the data? 30:30 - 31:00

Q9

do u know rag? 31:42 - 33:30 i ask: should i explain the detail response: nono...

Q10

a specific scene: The robot agent will generate a large amount of log data, which needs to be collected, formatted, processed, transformed, and then used to build a dataset for iterative model training. how to design the process.

my ask

summary

1. perform badly in the description of my project, for i speaking too much, and the interviewer do not interrput me. so later i have to concise my expression

2. perform badly in Q7 for i ignoring the redis + caffine part, this is because i prepare too much for the rag project, but rag is not presented in my resume, im sb

3. perform badly in Q10, for i do not remember what the interviewer said, and just retell what he said...

Q10 answer

collect logs -> kafak -> flink for data cleaning (empty session, trash logs, ineffective session)+ field normalization(session_id, agent_id, user_id, tool_input/output) + data masking(name, phone number) -> session_data used for sft, reward ,tool_input/output for agent / function call building ,vision data + action log for vla training. -> divide traing set and test set -> model launched and log re-streaming

i have to truly think what to do step by step and how to do(use what to implement) in the pipeline to solve the secene issue. and take a pen to record all the things the interviewer said

Back to Blog