Introduction

Here is my recap after finishing the Phase 1 in my Viettel Digital Talent 2026. It’s been several months since the first time I heard about this programme and decided to apply for it. Let’s go from the application to the first, biggest, solo, fast-paced project I had ever done in my undergraduate.

Prepare CV for application

One day, when I was doom scrolling on my facebook reels, I saw a program called Project X Vietnam 2026. Feeling some kind of curiosity, and also getting bored from a few months of not doing anything truly meaningful for myself, I decided to go straight to their fanpage. Then I saw a guy, who has the same major as me at university, on their poster with the name “Viettel Digital Talent” on his shirt. At that moment, I thought of perhaps I have just discovered something interesting enough to pull me out of this ultimate boredom so I go to their webpage to look for more information. Fortunately, the application is opening to the end of March 15, so I have a few days before it closes. The first round I need to pass is the TOEIC entrance test, which was totally predictable since I haven’t had any english certificates yet :D. The english test was easy for me though, if my memory serves me right, I got a 955 on the test, listening and reading only. On the same day taking the english test, I also had to participate in the IQ logic test, which included mysterious squares, rectangles and hidden patterns among them. The test is kind of easy, though I have a bad impression on these kinds of IQ test, since I strongly believe that they are totally useless to judge anything meaningfully about a human. Finally I got around 112 out of 160 points for the IQ test, average :v since some guys mysteriosuly got a full score of 160 or at least 140+, holy grails! After a few days, another entrance test appear, the aptitude test or “Bài kiểm tra chuyên môn”. This test is also the final one that I had to take, it will be different depending on which track you are participating, for example, I applied for track Data Science and Artificial Intelligence (DSAI) so I had to take the test related to these topics. Other track such as 5G, networking, digital design and verification, software engineering, data engineering, etc. would get a specially tailored test for them. Completed the final test, I got a 35/45 scores, which I considered not too high. Since the test contains some domain of data science (which is not my taste, my taste is deep learning) so I got nerfed in those questions, asking about SARIMA, ARIMA and time series processing. Although the score was not satisfactory for me, I heard that I has been passed straight to the program without an interview. Some guys on the forum said that only those with score around 33-34 may get an interview for further processing, and all guys with score >=35 all passed directly, whoops so lucky this time. And then, my journey with this program begins here.

First Lectures

At the beginning of the program, we only need to attend a few lectures each week. There would be lectures covering a wide range of topics: recommender system, speech processing, satellite image processing with spectral or something weird, finetuning large language models, transformers, MoE, and agentic systems. This stage is absolutely boring since we only sit there, open Zoom, watch the lecture or even mute it to do something else :V don’t be me. We were annouced that there wlll be a mini project for each of us but the details information was not fully available at this time. Therefore, I just skim over this stage and continue to the real challenge in Phase 1: the project. Although it were called mini, later I discovered that this project alone is on the same scale with some of my old projects in the past few semesters :D hooray such a surprise.

Project announcement

On the day the Mini Project was announced, there are around 122 projects in total. Okay, straightforwardly, it blew my mind. I didn’t think there would be so many projects, there were around 180 participants in the program so yeah each project contains only 2 slots to maintain the diversity as high as possible. Surfing around the list, I found some interesting and potential choice I could choose from. Currently I’m quite interested in generative models with diffusion/flow matching formulation and deep reinforcement learning (either policy training for general agents or reasoning training for LLMs). Therefore, I found some interesting projects worth my mind:

Self Evolving AI Agent for Telecom Failures

This project is indeed interesting from its name. The methods introduced by mentors of this project cover a wide range from touching the weights through finetuning with reinforcement learning to agentic-based methods such as memory management, skills and harnesses during agents lifecycle. Although sounds interesting, this project seems to me a software engineering projets rather than ML/AI project, since its tech stack is completely something. It covers a wide range of topics, including: AirFlow, Astronomer Cosmos, OpenLineage, Marquez, etc. Oh my god, so many mysterious terms! Because of this overwhelming tech stack, I decided not to go with this topic, despite its coolness ;D

Applying AI/ML in penetration testing and Redteaming

This project focuses on penetration testing and auto redteaming using agentic systems in network system management. Mainly the deliverables would be a CLI application that use an AI Agent to performing cryptographic attack for testing, and yeah cryptography is definitely the final thing I want to learn in this world. I have no knack for that field after a trial semester with it. Therefore, I didn’t go with this project either.

Adaptive Learning System

This project targetted at building a software for accelerating learning speed for high schoolers. The department in charged of this project is Viettel Business Solutions, so we can expect this project is probably the one that near the production stage the most in all these four projects I am talking about. Some of its features contains personalized syllabus, reduce redundant lessons, accelerate mastery speed, etc. Although this project mentioned some techniques related to reinforcement learning, I didn’t find it interesting enough to put my efforts into. Therefore, I didn’t go with this either.

Fucntion Calling using Reinforcement Learning with Verifiable Rewards

And here we comes with the final project I want to discuss among the four. This is the project that I chose to go with to conclude the project topic selection phase of the program. This project targetted at using RLVR to train a small language model to perform function calling (i.e, tool calling) tasks with a comparable performance with large language models. I chose this project mainly due to several reasons:

• It uses RL as core techniques for training, which is my favourite domain in deep learning these days, of course alongside with generative models. However there were no generative-related projects in the list, so I went with this.
• Its main contribution resolves around training and inference techniques, rather than building agentic system with a frozen model and a massive vendor lock-in that continuously tortures us with rate limits. As subjective as possible, I deemed projects that involve training and inference details AI projects, otherwise it would just be another software engineering project.
• There was another guy in the program that shares 2/3 of my name is also interested in this project, so I thought that would confuse the mentor a lot :D

Surprisingly enough, the mentor that guide me in this project was happened to be the champion of Viettel AI Race 2025, an AI competition hosted by Viettel themselves in 2025. As soon as I learned this, I feel kinds of lucky to pick just the right one to help me this project :D