Data Scientist Intern in Web3 Company: LLM Workflow & Blockchain
6 months ago, I started as a Data Scientist Intern at MEXC. I didn’t get a return offer because of the hiring freeze, but this internship has helped me a lot on practical AI implementation and helped me in my future job hunting journey.
Interview Process
The interview has 2 rounds, first round is HR interview, basically communicate to identify my information and set the interview time. The second interview is with hiring team, which included some tech problems, 1 python problem and 1 SQL problem, both are between easy to medium level, besides this, the team also asked me about how much I know about Web3. During the interview, I felt that having domain knowledge in the Web 3 space is quite important, if you want to step into this industry, learning some basic concepts of Web 3 will help a lot.
Projects at work
Over the last 6 months, my work varied between engineering LLM workflows and breaking down company data and blockchain data. Here is an overview of my projects at work.
Project 1: Intelligent News Automation
The Operations team needed to read daily crypto news. Between RSS feeds, competitor announcements, and market updates, they were spending hours manually filtering noise to find useful information.
My goal was to build an intelligent automation tool. Collaborating with data engineers who handled the ETL process to clean raw HTML, I focused on the brain of the operation via Dify.ai.
· The Hallucination Trap
My first attempt was a simple mistake. I tried to use a single instruction asking the model to classify and summarize news at the same time. The result is the model started “hallucinating” categories and missing key details. It was unreliable.
· The “Chain-of-Thought” Fix
To fix this, I re-engineered the workflow using a Chain-of-Thought approach. I split the task into two dependent steps:
- Agent A (The Classifier): Strictly tags the news with the correct topic.
- Agent B (The Summarizer): Takes the tag and the text to generate a concise summary.
While this split-process used more tokens, the accuracy rate is much more higher. This tool now generates the team’s daily morning briefings automatically, saving them about 1 hour of manual work every day. I later replicated this architecture to track competitor announcements, giving our business teams a real-time monitor on market shifts.
Project 2: Sentiment Analysis Pipeline
The Marketing team had a question about what the communities are saying about different tokens, they needed a “Social Listening” tool, but we didn’t have much budget for it. I initiated a project to analyze Twitter data via API. However, I hit a roadblock during the model selection phase.
· Cost vs. Performance
I initially planned to use commercial LLM APIs to analyze every single tweet. But I then found out that the expense for using APIs are completely out of our budget of this tool.
Get Xue Zhiming’s stories in your inbox
Join Medium for free to get updates from this writer.
· The Pivot
I had to pivot. Instead of expensive LLMs, I used local NLP models for basic sentiment scoring. It was a lesson in resource management: sometimes the best model isn’t the smartest one, but the most sustainable one.
I also faced a data gap: the API didn’t provide user location. To solve this, I engineered a Python script that inferred regions based on the language of the post. It was a heuristic approach, but it worked.
Project 3: User Analytics & CSAT
Beyond AI, a huge chunk of my work involved “Data Storytelling” using SQL and Dashboards. I investigated two critical issues: user churn and dropping Customer Satisfaction (CSAT) scores.
· For User Churn: I mapped the user journey funnel. By visualizing click-through rates, we pinpointed exactly where in the UI potential users were giving up. This was also a to-do list for the product team to fix the interface.
· For CSAT: I broke down support scores by region. This allowed managers to see exactly which markets were unhappy and adjust their support strategies accordingly.
Project 4: Blockchain Capital Flow
Perhaps the most “Web3” part of my internship was working with Dune Analytics. The Finance team needed to track large fund transfers. Blockchain allows for permissionless tracking, so I wrote scripts to track the capital flow and identify between private wallets and institutions. This experience reinforced why I love this industry: the transparency of on-chain data lowers information barrier in a way legacy finance couldn’t.
From Classroom to Practice
Reflecting on these six months, I can see a direct line between modules at SMU and my output during the internship.
- NLP for Smart Assistant & Generative AI: Understanding how tokenization works saved me when debugging my LLM prompts. I was adjusting the architecture based on first principles.
- Query Processing and Optimisation: 70% of my job was data retrieval, the SQL optimization techniques I learned were a lifesaver. Writing clean, efficient syntax meant I could answer business questions quickly.
- Data Science for Business: It taught me that a model is useless if you can’t explain it to a non-technical stakeholder.
Reflections
For the students looking for internships, I suggest you finish the module projects carefully, so that you can show your skills with the projects during interview. Also, remember to prepare with LeetCode before internship interview, you need at least to have the ability to solve medium level problems.
For me, this internship taught me that technical adaptability is valuable. Whether it was pivoting to local models to save costs or inventing a location proxy when data was missing, the ability to find a “Plan B” keeps projects moving.
Looking forward, I am convinced that the future lies at the intersection of AI and Blockchain. We are seeing the rise of protocols like x402 and Coinbase’s Agent Kit, AI agents can autonomously perform crypto transactions, and will perform more functionalities in the future.
Disclaimer: All opinions and interpretations are that of the writer, and not of MITB. I declare that I have full rights to use the contents published here, and nothing is plagiarized. I declare that this article is written by me and not with any generative AI tool such as ChatGPT. I declare that no data privacy policy is breached, and that any data associated with the contents here are obtained legitimately to the best of my knowledge. I agree not to make any changes without first seeking the editors’ approval. Any violations may lead to this article being retracted from the publication.
