Lessons From DevCon 2025
Reflections on Devcon25, from resilient Linux systems to efficient AI models and software testing with math.
A few days ago, I attended DevCon 2025, the 10th edition of the largest annual developers’ conference in Mauritius. The conference took place over the course of three days and featured nearly fifty national and international speakers. I first heard about DevCon two years ago but I was never able to attend it due to personal commitments. This year, I decided to make time for it despite being busy with my final year project. I attended seven talks in total during the first two days only.
In this post, I will be sharing my experience and I will rate each talk using the following highly subjective scale:
| Rating | Meaning |
|---|---|
| ⭐⭐⭐⭐⭐ | Excellent – Highly recommended, engaging and insightful. |
| ⭐⭐⭐⭐ | Good – Worth attending, solid takeaways. |
| ⭐⭐⭐ | Average – Fine but not memorable. |
| ⭐⭐ | Poor – Limited value, weak delivery or content. |
| ⭐ | Very Poor – Not worth the time. |
Talks I Attended
1. Building Unbreakable Linux Systems Through Strategic Failure
Speaker: Samuel Arogbonlo
Role: Senior Platform Engineer @ ChainSafe Systems
My Rating: ⭐⭐⭐⭐
Recommended Resources:
The first talk I attended was on chaos engineering, presented by Samuel, an infrastructure engineer. Chaos engineering is a testing technique where failures are deliberately added in the production or pre-production environment to analyze their impact (IBM, 2023).
Samuel emphasized the importance of chaos engineering through a series of real-life examples of companies who failed to use chaos engineering. One example he gave was when Slack’s Kibana cluster failed due to a lack of disk space. After fixing the issue, Slack implemented a Game Day during which its engineers deliberately broke things to see how the system would react. This activity helped uncover even more issues (Madden, 2024).
One simple way to use chaos engineering is to add artificial delays to your system to see what happens. Ideally, a system should be able to handle these delays without problem.
During question time, Samuel mentioned that it is usually the responsibility of the DevOps/infrastructure team rather than software developers to perform chaos engineering.
The talk was quite one-sided in the sense that chaos engineering was hailed as the holy grail to most problems. However, there are probably other factors (such as cost) at play which explain why teams are not actively using this technique. I wish these factors had been taken into consideration during the talk.
2. The Art of Transforming a Foundation Model into a Domain Expert
Speaker: Sebastien Stormacq
Role: Developer Experience Engineer @ AWS
My Rating: ⭐⭐⭐⭐
Recommended Resources:
In this talk, Sebastien gave a high-level overview of the transformer architecture. In short, an encoder converts a phrase to matrices/vectors (more commonly known as embeddings) while a decoder predicts the next token/s given a phrase. He then talks about five techniques for improving base AI models:
- Continued pretraining
- Fine-tuning
- Prompt engineering
- RAG
- Agents
He showed how to perform continued pre-training and fine-tuning using Amazon Bedrock but recommended against using a personal AWS account due to the expensive costs. During the talk, Sebastien also mentioned the Helpful, Honest, and Harmless (HHH) principle for evaluating base models.
During the question time, the speaker mentioned that hallucination can still occur irrespective of the technique you are using. He said that hallucinations are less likely with RAG and that sometimes it may be useful to use another model to validate the output of a model.
3. The Next Frontier: Exploring the Power of .NET Aspire
Speaker: Fawwaaz Koodruth
Role: Cloud Developer @ SDWorx
My Rating: ⭐⭐⭐
Recommended Resources:
- .NET Aspire
- meetup.mu
Even though I had zero knowledge of .NET and this talk not meant for beginners, I decided to attend it with an open-mind and I actually did learn a few basic things:
- .NET Aspire makes deployment of .NET applications easier by avoiding complex Docker and Kubernetes YAML configurations.
- NuGet is the package manager for .NET.
- The OpenTelemetry Standard provides a framework for collecting telemetry data (logs, metrics, traces).
- Podman is a Docker alternative.
- There is a local user group for cloud enthusiasts called .NET User Group of Mauritius.
4. Beyond LLMs: The Rise of Small, Efficient AI Models
Speaker: Kavish Seechurn
Role: Tech Lead @ MCB
My Rating: ⭐⭐⭐
The whole talk was light-hearted and it was very approachable for beginners. First, Kavish explained the difference between an LLM and smaller language models in layman’s terms using analogies. He also took into account the number of parameters, MMLU, and latency. He then demonstrated how to run a tiny AI model on Google Colab before explaining how to choose an AI model for local use based on use case.
However, I was quite disappointed when Kavish mentioned that he had no experience running AI models locally using tools like Ollama. I think that the whole point of using small AI models is to be able to run them locally. Using Google Colab to run small AI models is not a fair comparison because it has more VRAM and a more powerful GPU than the average consumer grade hardware. To be fair, Kavish acknowledged that he would have preferred to give a talk on a topic in which he had more experience, but unfortunately, his proposals in his areas of expertise were not selected.
5. Exploring OpenData Mauritius with Python Marimo
Speaker: Kherin Bundhoo
Role: AI Engineer @ Telesio France
My Rating: ⭐⭐⭐⭐⭐
Recommended Resources:
In this talk, I learned the three main issues with Jupyter Notebooks that Python Marimo solves:
| Feature | Jupyter Notebook | Marimo |
|---|---|---|
| Reactivity | Not supported | Fully reactive UI updates |
| Built-in UI widgets | Not available natively | Yes – Comes with UI widgets |
| Version control friendliness | Hard to diff/merge (stored as JSON) | Easy – Standard Python code format |
What I particularly liked about the presentation was that Kherin first demonstrated the pain points of using Jupyter Notebook in a real-life example where he was performing some data analysis on the times in Dodo Trail. Then, when he introduced Marimo, everything just clicked and made sense to me. I will definitely be using Marimo for my next data analysis project.
This talk also had a notable engagement from the audience, with a lot of questions towards the end. When asked about the long term viability of this project, Kherin said that the project is open-source and is used by developers at Fortune 500 companies so it is unlikely for the project become unmaintained in the future. However, I must add that Marimo does not have the same community support as Jupyter Notebook. For example, Google Colab only supports Jupyter Notebook although there are alternative options:
- Use molab instead of Google Colab. Molab as created by the creators of marimo and offers similar compute and sharing capabilities.
- Convert Marimo to Jupyter and then upload it to Google Colab.
Kherin also talked about the OpenData initiative of the Mauritian government whereby datasets on many areas like health, education, transport, and much more were publicly available. One caveat of Mauritius OpenData is that some datasets are not always kept up-to-date with cutoff dates around 2018.
6. Exploring the openSUSE Project
Speaker: Ish Sookun
Role: Systems Architect @ La Sentinelle
My Rating: ⭐⭐⭐⭐
Recommended Resources:
- openSUSE
- openQA
- Curated openSUSE resources
This talk was another surprise talk that I attended with no prior knowledge of the topic of discussion. It was interesting to learn the history of openSUSE and how Mauritians are making open-source contributions to it. Ish also hinted at an upcoming project made by Mauritians for openSUSE. One funny thing I noticed was that Ish, an openSUSE board member, gave his presentation on macOS while praising openSUSE.
Some of the new things that I learned:
- MIXP (Mauritius Internet Exchange Point) is the first internet exchange point in Mauritius. It allows peering members (e.g. Telecom, Emtel) to route traffic between each other in Mauritius. This avoids traffic having to go through international connections, reducing latency.
- top500.org is a website that maintains a list of the 500 most powerful computers in the world. All the supercomputers are running Linux.
- There are two distros of openSUSE: Tumbleweed and Leap.
- Slackware is the oldest Linux distro.
- OpenQA is a testing tool for operating systems.
7. Where Mathematics Meets Software Testing
Speaker: Lovelesh Beeharry
Role: President of Mauritian Software Testing Qualifications Board (MSTQB)
My Rating: ⭐⭐⭐⭐
Knowledge of mathematics can be useful in many fields, including software testing. Here are some applications of mathematics in software testing:
- Formal logic can be used to define test cases.
- In pairwise testing, the combination formula
nCrcan be used to calculate the number of test cases needed. A good explanation is available in this blog post by NashTech Global. - Statistics: Mean, mode, median, standard deviation, and cumulative frequency curve are often used in performance testing.
- Differentiation: Given a graph of response time against concurrent virtual users (CVU), differentiation can be used to analyze the rate of change of response time as load increases. This helps identify performance tipping points, where response time begins to degrade rapidly.
- Code optimization: Knowledge of space and time complexity is essential. Time complexity models how the execution time of an algorithm grows with input size, while space complexity models how much memory the algorithm requires as the input size increases, especially as it approaches infinity. Understanding both helps developers write more efficient, scalable code.
There were probably a few more applications that I missed because I came late to the talk. Overall, the talk was quite interesting because I was able to relate to some of the applications mentioned.
A surprising claim from Lovelesh was that international companies like Cloudflare and Accredible were avoiding expanding into Africa due to a lack of ISTQB accreditation.
Honest Critiques
I want to preface this section by acknowledging that Devcon25 faced financial difficulties this year, which undoubtedly impacted the scale and logistics of the event. With that said, I had high expectations for this conference since Google and Microsoft were Platinum (highest tier) sponsors. Unfortunately, my expectations fell short:
- There were basically no activities, other than a memory game from SWAN, to do in between the talk sessions. I expected interactive games and creative challenges as stated on the MSCC website.
- The rooms were quite small and for some sessions like the panel discussion and the unbreakable linux session, some attendees had to stand. I arrived late to the panel discussion and stood for around 10 minutes before leaving to attend another session. One thing I appreciated with the small rooms was that it was much easier for the audience and the speaker to interact without a microphone.
- No free food or drinks were provided to attendees, not even a snack.
- The initial goodies provided upon registration were underwhelming. I did not even receive a notebook and a pen but the lanyard by cloud.mu was cool though. Also, credit to the speakers in the last two talks I attended, as they actually handed out some goodies during their quiz sessions.
- There were not enough booths from sponsors. I wish each sponsor had set up a booth so I could learn more about them.
Conclusion
I am grateful for the opportunity to attend such a conference for free. Despite the tighter budget, but I think that the organizers still did a decent job. I was able to expand my knowledge on a wide range of topics and tools and I am definitely looking forward to DevCon 2026!
References
- Madden, S., 2024. Break Stuff on Purpose [online]. Available at: https://slack.engineering/break-stuff-on-purpose/ [Accessed 26 July 2025].
- IBM, 2023. What is chaos engineering? [online]. Available at: https://www.ibm.com/think/topics/chaos-engineering [Accessed 26 July 2025].