Meet Your AI Data Scientist!

A test drive with Colab + Gemini 3 Pro

Why am I doing this?

I have a paid subscription to Colab Pro and decide to test it as my co-Data Scientist. My subscription includes

• 100 GPU and TPU compute units
• coding assistance powered by Gemini (2.5 Flash and 3.0 Pro)
• access to a series of Gemini services without using an API key.

Some background information on the tester: I am a researcher in an applied field. Data analysis and prediction are essential to my work. I am fluent in Python, though coding has not been my focus in recent years.

The test: “Modelling **Singapore dollar Nominal Effective Exchange Rate (S$NEER)**”. This subject is chosen based on the following considerations:

• It is a problem that requires training in both Macroeconomics and central banking practices.
• I have broad exposure in this area but am neither an expert nor a practitioner in this subject. In another word I know enough to cross-check what Gemini comes up with, meanwhile I would also keenly appreciate if Gemini can help deepen my understanding.
• It requires data science as well as coding skills

And here are the lessons learned

Does Gemini know what data to get and retrieve correctly?

Mostly. Gemini correctly identifies Monetary Authority of Singapore (https://www.mas.gov.sg/) and Singapore Department of Statistics (https://www.singstat.gov.sg/) as two key data sources. It is also able to come up with the code to retrieve and parse the data. However it fails when the site is mostly designed for a human developer. For a few times I have to either find the right API information (buried in dynamic contents) or eyeball the json files myself and then instruct Gemini what to do next.

Does Gemini actively do data quality checking without being prompted explicitly?

No. You need to take data QA 100% into your hands. For example, to verify the data downloaded, I have to cross check them with other data sources and do find missing data points due to a bug in code. Of course if I have prompted Gemini, it would have come with some ways to check. But without instructions it would not actively pursue it. Same lessons for identifying erroneous data and outliers. It knows how to do it. But it won’t do it unless prompted to do so.

Is Gemini able to come up with sensible modeling approaches?

Yes. This is the part of the test I find most satisfactory. Gemini, for safety probably choose to use the best-documented or the most popular models. If your goal is to find new modelling approaches, you can prompt Gemini to do research for you. Again you need to be proactive and quizzical.

It can code and debug its own code quite well. With Gemini we also do not need to write plotting code anymore.

Can I use AI models inside my data science project in Colab?

Yes. I use AI to analyze the MAS Policy release for S$NEER modeling . Currently in early 2026 there are three ways to access AI models in Google Colab:

Load AI models into your own virtual environment on GPU or TPU. This is the way to go if you want to do AI oriented research. If you want to do prompting with commercial AI products, you should use one of the two following approaches:

Acquire your own API key and access Gemini, Claud, or other AI products.

Without API key, use google.colab.ai to directly access Google AI models. Note you have to ask the Code Assistant explicitly for this. Otherwise it will by default go for the API key approach.

Can Gemini interpret data analysis and modeling results independently?

Not really. I would not trust AI to do this alone as of January 2026.

In the first case I ask it to describe how Singapore’s trading partners trade weights change with time. Given data and plots this is an easy task. Yet What Gemini describes does not match the data plot in question. It sounds more like some analyses it has learned online.

In the second case I ask it to explain the massive reserve draw-down in 2022. This is a harder task, as the information does not lie in the data tables in the project, but somewhere online. Curiously, Gemini (the coding assistant) knows the fund transfer to GIC in 2022, but it some how fails to link the two facts together. It insists that the massive draw down is due to MAS intervention in the market. To give Gemini credit where it is due, when I ask my Gemini 3.0 Pro chatbot the same question without any data analysis involved, it correctly states that the drawdown is due to the fund transfer to GIC. It is possible that the data analysis changes how Gemini synthesizes context information during inference time.

Conclusion

In summary I find Gemini inside Colab extremely helpful and here are the things might be useful if you are going to supervise an AI data scientist:

• You can give a detailed research plan if you have one. You can also just ask broadly. Either way you need to participate actively in the research step by step to ensure the quality of the search.
• Do not use the “accept and run all” option unless it is a very small task such as plotting. For more complex tasks Gemini times out on me quite often.
• Clean up the old cells if they are not current anymore, especially those cells with planning information. Gemini uses texts in cells as part of the context. If you do not delete, Gemini may say what it is going to do next, but end up doing the old thing.
• Leave almost all the coding job to Gemini but review the code focusing on the key logic.
• Gemini is a very knowledgeable coworker to bounce ideas off. Prompt strategically to get the most out of it.
• Pay close attention to the interface between data and the world. This includes data quality control, data interpretation and any other steps in the pipeline where metadata, world model, and domain knowledge are key to the integrity of the work. If uncertain, use online searching or a second AI product to cross check what the coding assistant is telling you.
• Happy Data Surfing!