No code? No problem: Andrey Gavrilenko shows how to break into Data Science with just English and an LLM
Break into data science with just English and an LLM. Andrey Gavrilenko shares how, and why responsibility still matters.
In a session that raised eyebrows and sparked plenty of Slack DMs, technical speaker Andrey Gavrilenko laid out a compelling,and surprisingly practical, case for becoming a data scientist without writing a single line of code.
His argument?
You don’t need to be a Python wizard anymore.
All you need is English and access to a solid large language model.
The talk, delivered as part of a developer-focused AI summit, raised a question that’s top-of-mind for a growing segment of tech newcomers and non-technical founders: Can LLMs replace traditional code workflows for data science?
Andrey says yes, at least to get started.
AI-powered Data Science – no code required
Ten years ago, Andrey got his start in data science through an Andrew Ng course. Now, he says, we’re living in a world where AI is becoming as fundamental as electricity.
And, for those who understand how to ask the right questions, that world no longer requires code.
LLMs like ChatGPT and Claude are unlocking the ability to perform complex data workflows with nothing more than natural language.
That means anyone, from startup founders and PMs to high school students, can run exploratory data analysis, build MVPs, and surface insights, all by typing what they want.
But this new frontier has its limits. Prompts can be misinterpreted. Results may vary. And if you’re piping sensitive data through a public API, there are real risks.
For larger datasets, there’s also the constraint of limited memory – most models can’t process huge volumes of information at once.
Slide deck from Andrey’s presentation covering why less coding is required for Data Science.
Still, Andrey’s message was clear:
“This changes who gets to build. It’s no longer about knowing Python – it’s about knowing what to ask.”
From Web Summit to real-world insights
To prove his point, Andrey walked the audience through an analysis of a real dataset from Web Summit 2023 in Lisbon.
It included information on over 2,700 startups, including descriptions, categories, and growth stages.
screenshot from Andrey’s presentation showcasing records from Web Summit startups
Using ChatGPT, he:
Cleaned duplicates and missing values
Analyzed stage distribution (Alpha, Beta, Growth)
Surfaced top categories like AI and SaaS
Ran comparisons across countries
Explored pitch deck availability
Broke down support partners
All of it, visuals included, was generated with prompts, not scripts.
What would typically take hours of Python took minutes in natural language.
It wasn’t flawless, at one point, the model miscounted country totals but it worked well enough to validate the approach.
Getting even deeper: category extraction and web scraping
Andrey didn’t stop at basic EDA. He asked the model to infer multiple categories per startup, based on descriptions.
The results were mixed: the LLM surfaced secondary categories but lacked consistency across the full dataset.
His fix?
Break the task into smaller chunks or offload it to a locally run script.
Naturally, the model offered to write that script too, automatically generating Python code to scrape metadata from startup websites.
In under two minutes, it returned enriched data from five sites.
How it actually works under the hood
To unpack the architecture behind this magic, Andrey walked through the role of the Model Context Protocol (MCP), a growing standard in AI infrastructure.
MCP connects AI agents to tools like calendars, databases, and design platforms using a secure, shared protocol.
In this flow, the user gives a prompt, the AI agent interprets it, and MCP bridges the gap to the actual tools needed to execute the task.
It’s a way to plug natural language into real-world software systems.
Vibe coding: the cool and the caution
Andrey acknowledged the hype around “vibe coding” – the idea of building software by prompting AI tools rather than writing code.
While the concept was popularized by former Tesla AI director Andrej Karpathy, Andrey pushed for a more responsible approach.
“There’s a dark side,” he said. “Lost work, security gaps, rogue agents making unintended changes. You can’t just fire off prompts and hope for the best.”
His advice?
Follow a few best practices:
Use version control to catch changes
Ask your agent to explain itself
Use trusted MCP servers only
Choose LLMs that are reliable and benchmarked
Create rules and use hint files to guide behavior
One feature Andrey highlighted was Goose’s .gooseignore – a file that blocks the AI from accessing sensitive areas, like test environments or secret keys.
AI as teammate, not replacement
Andrey closed with a message of balance. Language models are changing the game but they’re not magic. You still need core knowledge of programming, data theory, and how systems work.
Tools like Pieces for Developers, he noted, are helping with long-term context management, enabling more advanced workflows while still grounding users in technical fundamentals.
“We’re in the early innings,” he said.
“But the opportunity is massive. Just by knowing English and having access to a strong LLM, you can do a lot. The next step is learning how to build safely and smartly.”
