Written for confused recent graduates: how to win more opportunities in the AI era

When preparing to find a job, it's easy to spend all our time on preparation. We pile up courses, always feeling that we need to learn a little more before we have the confidence to submit resumes. Now recruitment has added RAG, Agent, and full-stack, so the learning list keeps growing longer, and people become more confused.

Some big-company internship positions have already put model integration, RAG, Agent, tool calling, and evaluation into their requirements. When we do projects, we also need to think about how these capabilities apply to specific business scenarios.

For students preparing to apply for AI application, Agent, or AI full-stack roles, I suggest upgrading your main project in these directions. Also, find an internship as early as possible, and open source and public writing can be done around the project.

1. Prepare directly according to the target position

First, I find about ten or so internship or campus-hire positions at big companies that I want to apply to, separate the common requirements from the bonus points, and then look at recent interview experience for the same direction and same graduation year. Arranging learning based on this information is more useful than chasing every hot topic.

Full-stack can have a main direction, but you need to be able to connect pages, interfaces, databases, and deployment. For AI applications, you also need to be able to explain where the model is involved, how data flows, and how to troubleshoot errors.

Keep learning the basics. When preparing for AI application or full-stack roles, besides RAG chunking, retrieval, and evaluation, also prepare thread pools, language fundamentals, and algorithms. For any knowledge used in a project, go back and understand the principles and applicable scenarios.

After each interview, write down what you didn't know, or what you half-understood and couldn't explain clearly. This gives you focus for the next round of preparation.

2. Make RAG or Agent something you can explain and verify

If these roles are your goal, an admin dashboard or a chat page that connects a model can be a starting point, but the main project needs to go further.

I suggest choosing a specific scenario, first getting the basic flow working, then going deeper into one or two problems. The two routes below: pick one as your main project.

For RAG, start with document Q&A that provides sources.

Build a knowledge base from public manuals or course materials, and make sure answers can find sources. When material is insufficient, state that clearly. After the basic flow works, continue handling the following problems.

How to process materials: How to split long documents and tables, how to preserve original positions, and how to make old content exit retrieval after updates.

Why answers are wrong: Whether it's no material found, unrelated content retrieved, or having the basis but still answering wrong, you should be able to distinguish from the records.

Whether changes are effective: Prepare a batch of questions you have manually verified answers to, covering different phrasings, combining multiple sources, and cases with no answer in the materials. Use a fixed set of questions to compare the effect, time, and cost before and after changes, and keep another set that did not participate in debugging for final verification.

First build a baseline version that can be compared, then consider hybrid retrieval, reranking, and other solutions based on failure cases. You need to be able to explain the splitting basis, how to sync after updates, and how to annotate evaluation data. Also report sample size, test conditions, and known limitations of how good the effect is.

For Agent, choose a task that requires calling tools to complete.

For example, a ticket assistant: based on user description, look up documents, query existing tickets, create a ticket after confirmation, then show processing status. Using mock data and two or three tools, you can start.

Then go deeper into one or two problems: what to do when tool arguments are wrong? After an interface timeout, will retrying create duplicate tickets? After a task is interrupted, can it resume execution?

Keep call logs, set execution count or cost limits. For write operations, prevent duplication; after a timeout, first query the actual result. Finally, verify whether the task was completed through the database ticket records. This is also the actual result that Agent evaluation needs to pay attention to.

AI programming tools can help, but you must read, modify, and verify the key code yourself, be able to explain the design, and continue to change requirements.

Finally, organize the repository, run instructions, demo, and evaluation results, and find a few people to try it out. In your resume, clearly write what you changed and how you verified it. Also honestly state if the sample is small and the effect is limited. A project that can clearly explain its trade-offs and failure reasons is one worth continuing to discuss.

3. If you can do an internship, start as early as possible

This is something I would put very high priority: if you can do an internship, start as early as possible. If school and time allow, start applying. Don't wait until all the above projects are done.

Work on projects while looking for internships. Don't only look for job titles with "AI." Opportunities to participate in real development, have code reviewed, follow through on requirements, and go live are all worth serious consideration.

In a company, modifying existing systems, integrating models, and handling failed requests all require collaboration with others. Getting exposed to this work early helps you discover problems that solo projects never surface.

During interviews, ask clearly what the actual work is, whether someone will mentor you, and which part you can participate in. Big companies are fine, and companies with suitable business and teams are also fine. If graduation is near, push campus recruitment and formal positions at the same time. Don't pin all hope on the next internship.

4. Start open source from the tools you are already using

If you get the chance to participate in open source, I suggest you try it, and ideally start with tools you have actually used in your projects.

When a parser fails to process a certain file, a tool-calling example is outdated, or an evaluation script lacks an edge case, you can first reproduce it, then read the contribution guide and discuss with maintainers. After confirming the problem, submit a PR, add tests, explain the change, and follow through the review.

Such contributions leave a record of code, issue analysis, and collaboration. Put the relevant links in your resume; they also give you concrete content to talk about in interviews. By continuing to participate in a project you're interested in, you gradually meet developers in the same direction.

5. Write up your project experience, and maybe make a podcast

I suggest finding a place early to publicly record your learning and project experience. You can write articles on Juejin, Zhihu, or WeChat public accounts, or start your own blog or podcast to talk about projects you've done, pitfalls you've hit, and your understanding of technology.

Take content from real development. Why does the RAG fail to search when the answer exists in the materials? After adding reranking, what changes in effect and time? Why did the Agent create duplicate tickets? Explaining how the problem appeared, how you diagnosed it, and how you solved it is content worth sharing.

The podcast doesn't need to be complex at the beginning. A dozen minutes to explain one problem, then put code, demo, and related materials in the show notes, so interested people can learn more.

If you are a team leader or company owner, after reading my article and listening to my podcast, if you feel my abilities and way of working fit your team's needs, are you more likely to proactively come ask me whether I'm looking for a job, instead of contacting other people who are also reading similar content?

So, I put the projects, articles, and shows that best demonstrate my abilities into my resume, and also leave convenient contact information. First choose a channel you can sustain, let others see what you've done and how you think, and that becomes another road to opportunities.

6. Summary

If you're still in school, I suggest doing three things this week: organize target positions, start applying for internships, and add a set of repeatable verification questions or tasks to your main project. After each problem you solve, update the code and records. When a suitable open source issue appears, get involved.

If you're already in the graduation job-hunting stage, put applications, written tests, and interviews first, and fill in projects, open source, and content output as time allows. Preparation and job hunting can happen at the same time. Don't let an ever-growing learning list keep pushing back the time you submit your resume.