I help founders and leaders build AI products that earn trust, scale responsibly, and keep humans in the loop.

AI, which I call Synthetic Agency (SA), is a great technology in the right context, is a harmful technology in the wrong context, and is a low ROI feature in most contexts.

I've built synthetic agency in products from the ground up, launched ventures in the field, performed academic research and contributed to think tanks, wrote The Ethics of AI, and advised 100+ leaders over the past 12 years.
I’ve learned that, while most people attribute failure modes to technicalities or talent, the true failure is philosophical.

The products that succeeded long term had teams who did not skip the hard questions. They built synthetic agency when they had to, not because they just could. They solved real problems by listening to their customers without chasing trends.

If you're working on AI, with AI or are figuring out how it impacts your work, and want to:
• Build something people actually want to use 
• Think through the ethical stuff without getting paralyzed
• Understand how your brand discovery and promotion should change with AI engines in town

I might be able to help.

I also write for 15,000 people on LinkedIn and for 3,000 on my newsletter about honest tips to make sense of this technological shift.

I'm also launching a monthly private newsletter. Invite-only. For Founders, CMOs, and Heads of Growth who need to control their brand voice in AI conversations (before it becomes too late or too expensive). DM me if you want to get an invite.

🎤 Open to speaking, podcasts, and collaborations—DM me or visit albertochierici.com

Alberto Chierici

I am a Data Scientist fueled by curiosity and a commitment to continuous learning, constantly seeking new opportunities to explore and grow.

My experience spans diverse industries, working with a wide range of data, all tied together by a shared drive to tackle complex and intriguing challenges.

I am particularly passionate about Natural Language Processing, whether it's delving into text classification, sentiment analysis, or language generation.

💬 Are you interested in the same topics or just want to discuss more? Let's connect, my inbox is always open for discussion!

🔍 #DataScience  #NLP #Innovation #AIEnthusiast #DataDrivenInsights

Simone Fuscone

The ultimate challenge: 

explaining the normal distribution without saying a word.

Time to head to the gym and put your stats knowledge to work.

Talking about Data/AI - Data Scientist | Ph.D.

Data Scientist before 2020:

I used my math and statistical backgrounds to achieve a 5% improvement with this model that can run with 256Mb memory

Data Scientists after 2020:
I’m using a 32B LLM to solve a linear classification problem


What do you think? 

***

Follow Simone Fuscone to help me create more content
Subscribe to get notified when I post 
Share the post with your network

6 Concepts You Need to Know Before You Start Studying Statistics

[for any data profile]

Statistics can feel overwhelming at first.

But mastering just six core concepts will give you a solid foundation.

If you're analyzing business data, these building blocks will help you understand patterns and make informed decisions.

Here are the 6 concepts:

1. Descriptive Statistics
This is necessary to summarize and describing the features of a dataset. 
It uses measures of central tendency (mean, median, mode) and variability (range, variance, standard deviation) to provide a "snapshot" of what your data looks like without making any broad conclusions.

2. Probability Distributions
Understanding distributions like Normal (Gaussian), Binomial, and Poisson helps you predict how data points are likely to behave and identify outliers that don't fit the pattern.

3. Hypothesis Testing
This is the formal process of using statistics to determine whether a specific claim about a population is likely to be true. By calculating a p-value, you can decide whether your results are "statistically significant" or if they likely happened by random chance.

4. Regression Analysis
Regression is used to model the relationship between a dependent variable and one or more independent variables. It’s essential for predictive modeling, allowing you to estimate how a change in one factor (like marketing spend) affects another (like total sales).

5. Statistical Significance and p-values 
While related to hypothesis testing, the concept of significance is a pillar on its own. It provides a threshold (usually alpha = 0.05) to help researchers minimize Type I errors (false positives). It ensures that the patterns you see in your data are robust enough to base decisions on.

6. Central Limit Theorem (CLT) The CLT is the "backbone" of statistics. This allows us to apply powerful statistical techniques even when our data is messy or non-normal.


***
Follow Simone Fuscone
to help me create more content
Subscribe to get notified when I post
Share the post with your network

You can now run 70B LLMs on a 4GB GPU.

AirLLM just made massive models usable on low-memory hardware.

how?

Have a look at AirLLM.

AirLLM released memory-optimized inference for large language models.

It runs 70B models on 4GB VRAM. It can even run 405B Llama 3.1 on 8GB VRAM.

AirLLM loads models one layer at a time.

Instead of loading everything:

1/ Load a layer 
2/ Run computation 
3/ Free memory 
4/ Load the next layer

This keeps GPU memory usage extremely low.

𝗞𝗲𝘆 𝗱𝗲𝘁𝗮𝗶𝗹𝘀 
• No quantization required by default 
• Optional 4-bit or 8-bit weight compression 
• Same API as Hugging Face Transformers 
• Supports CPU and GPU inference 
• Works on Linux and macOS Apple Silicon

You can run several open-source models; clearly, inference times can be slow.

***

Follow Simone Fuscone to help me create more content
Subscribe to get notified when I post 
Share the post with your network

The difference between a RAG system that hallucinates and one that doesn't?

It starts with how you chunk your documents.

In Retrieval-Augmented Generation (RAG), chunking is the process of breaking down large documents into smaller, manageable pieces.

This is crucial because embedding models have token limits, and smaller chunks help the system retrieve only the most relevant information rather than a 50 page PDF.

Here are the primary methods used to chunk documents:

1/ Fixed-Size Chunking:
The most basic method. 
You decide on a specific number of tokens or characters per chunk (e.g., 500 tokens) and split the text regardless of content.

How it works: Text is sliced at a set interval. Usually, an overlap (e.g., 10%) is included so that context from the end of Chunk A is repeated at the start of Chunk B.

Pros: Extremely fast and easy to implement.
Cons: Often breaks sentences or paragraphs in the middle, losing semantic meaning.

2/ Recursive Character Splitting:
The "industry standard" for many RAG applications. 
It attempts to split text based on a hierarchy of characters to keep related content together.

How it works: It tries to split by double newlines (paragraphs). If a chunk is still too big, it tries single newlines, then spaces, and finally individual characters.

Pros: Keeps paragraphs and sentences intact as much as possible, preserving context.
Cons: Still "blind" to the actual meaning of the words; it's just following formatting rules.

3/ Document-Based Splitting
This method respects the inherent structure of the file type (Markdown, HTML, or JSON).
How it works: It uses the document's syntax to define boundaries. 
For example, a Markdown splitter will create new chunks at every Heading or Subheading.

Pros: Highly organized; maintains the relationship between headers and their content.
Cons: Only works if the source document is well-formatted and consistent.

4/ Semantic Chunking
A more advanced approach that uses AI to determine where one idea ends and another begins.

How it works: The system calculates the "distance" (similarity) between sentences using embeddings. When the meaning shifts significantly between Sentence A and Sentence B, a split is created.

Pros: Captures complete ideas and themes perfectly.
Cons: Much slower and more expensive, as it requires running an embedding model just to perform the split.

Which one have you used? 

***

Follow Simone Fuscone to help me create more content
Subscribe to get notified when I post 
Share the post with your network

Would you like to land an interview without experience?

Stop using Kaggle datasets like we were in 2017.

You need to build a portfolio that spans Data Science (DS), Data Engineering (DE), and AI Engineering (AI).

It requires projects that demonstrate how data is transformed from a raw state to a deployed, intelligent application.

Using a notebook to build a POC is no longer enough.

Here are 4 examples of end-to-end projects you can build:

1/ Cross-Platform Sentiment & Trend Engine

The Idea: A "Brand Intelligence" tool that scrapes Reddit, X (Twitter), and YouTube comments daily to map how sentiment for a brand (e.g., "Tesla") is shifting compared to competitors.

Technical Stack:Scrapers: Playwright/Apify/RapidAPI.
NLP: Hugging Face Transformers (Emotion classification).
Dashboard: Tableau or PowerBI connected to a Postgres database.

Workflow: Build a daily scraping cron job. Run sentiment analysis on every comment. Aggregate scores by date and platform. Deploy the pipeline on a Cloud Run (GCP) instance.

2/ Real-Time Sports Odds & Outcome Predictor

The Idea: A live dashboard that ingests real-time game stats (e.g., NBA or Premier League) to predict win probabilities and "value bets" as the game progresses.

Technical Stack:
Data: Sports APIs (RapidAPI)
Storage: DuckDB (local analytics) or Snowflake (cloud).
ML: XGBoost for tabular data, Optuna for hyperparameter tuning.
Deployment: FastAPI for the prediction engine, Streamlit for the frontend.

Workflow: pull API data. Train a model on historical seasonal data. Deploy a consumer that feeds live data into the model and updates a dashboard.

3/ Legal-Compliance "Active" RAG System

The Idea: A tool for legal teams to upload 500+ page contracts and ask complex questions like, "Where does this conflict with GDPR 2025 updates?" Unlike basic RAG, this uses Agentic Retrieval to reason across multiple documents.

Technical Stack:
LLM: Claude Sonnet/GPT/Gemini.
Orchestration: LangGraph or CrewAI (to handle multi-step reasoning).
Vector DB: Pinecone or Weaviate.
Processing: Unstructured (for complex PDF parsing).

Workflow: Create a pipeline that chunks documents based on semantic structure. Use an agentic flow where the AI "searches, reflects, and re-searches" if the initial answer is insufficient. Deploy via Docker on AWS App Runner.

4/ Personalized Health Metric "Forecaster"

The Idea: A multimodal app where a user uploads a photo of their meal. The AI calculates the caloric impact

Technical Stack
Vision: LLaVA (Vision-Language Model).
Time-Series: LSTM or PatchTST.
App Framework: FastAPI + React Native.

Workflow: Use a Vision model to identify food from images. Merge that "textual" info with numerical wearable data. Deploy

Which one would you choose?

***

Follow Simone Fuscone to help me to create more content
Subscribe to get notified when I post 
Share the post with your network

Your data has a story …

are you telling it right?

You built the model. 
You ran the analysis. 
Now comes the hard part:

proving it matters fot the business case.

Tools like Flourish help you improve this process, enabling beautiful, interactive visualizations that captivate stakeholders.

It makes your message instantly clearer.

Key aspects:
1/ Converts complexity into clarity for any audience
2/ Engages decision-makers with memorable visual impact
3/ Accelerates understanding and speeds up decisions
4/ Transforms passive viewers into active participants

if you think it’s a kind of social media approach to get attention, you are right. 

Because it works.

link in the comments

***

Repost, save or share with someone who needs to improve data storytelling
Follow Simone Fuscone

Documents weren't built for machines.

This is how you can feed your AI.

For decades, we've created files for human eyes, 

PDFs for reading,

Excel for analyzing

PowerPoint for presenting

But now we need our AI systems to understand this information too.

That's where the gap appears: machines don't naturally comprehend the formats we've been using all along.

MarkItDown transforms diverse files into clean Markdown.

PDFs, PowerPoint, Excel, audio, images. all become AI-ready text.
Not just any text.
Structured, meaningful Markdown.

Why use it?

Markdown is the language AI models speak fluently.

Native understanding, models trained on vast Markdown datasets comprehend its structure naturally

1/ Preserves context - Headings, lists, tables maintain information relationships
2/ Token efficiency - Minimal markup means lower costs, faster processing
3/ Universal compatibility - Works across all major language models
4/ Human-readable - Audit and verify what your AI receives

We have a huge amount of data sitting in our repositories. 

Business reports in slides. 
Financial data in spreadsheets. 
Research in PDFs.

MarkItDown makes it easier to use them. It converts existing data into AI-ready format instantly.

***

Follow Simone Fuscone to help me to create more content
Subscribe to get notified when I post 
Share the post with your network

Theory won't get you hired.

Stop buying expensive courses and start building

The path is clear: 

Companies want builders, not just POC makers.

AI can make your ML pipeline in 3 seconds.

You need to show that you can build something that is reliable, and then you can maintain.

This FREE repo contains everything you need:

OCR & Vision - Extract text, process images, build visual AI
Chat Interfaces & UI - Local ChatGPT clones with custom models
RAG Systems - From basic retrieval to enterprise-grade pipelines
AI Agents & Workflows - Multi-agent systems that actually work
Voice & Audio - Real-time conversational AI
Multimodal AI - Combine text, images, video, and audio
MCP (Model Context Protocol) - Connect AI to your tools
Fine-tuning & Production - Deploy real systems, not demos

divided into 3 difficult levels: beginner, intermediate, and advanced

Link in the comments

***

Follow Simone Fuscone to help me create more content
Subscribe to get notified when I post 
Share the post with your network

The common thread among elite data engineers?

It's surprisingly boring.

The best data professionals aren't the ones building the fanciest models. 

They're the ones solving actual problems with solutions that won't break at 3 AM. 

In data science and engineering, we often overemphasize complexity. 

We chase the latest framework, the newest architecture, the most exotic algorithm.

But we forgot that "Simplicity is a prerequisite for reliability." 

Not a suggestion. A prerequisite.

Consider the data engineer who chose CSV files and cron jobs over a complex streaming platform. 
The pipeline ran for three years without incident. 
No streaming state to debug. 
Just boring, bulletproof automation. 

Or the data scientist who deployed a logistic regression instead of a deep learning model. The business got predictions they could explain, maintain, and trust. 
The model scored 82% accuracy. 
The gradient boosting alternative? 84%. 
The neural network? 
Also 85%, plus three weeks of hyperparameter tuning and a GPU budget. 

Which solution actually solved the problem?

The sophisticated data professional asks different questions. 

1/ Will this run reliably in six months? 
2/ Can someone else maintain this? 
3/ Does this actually solve the business problem? 
4/ What breaks when I'm on vacation?

The market doesn't need more data professionals who can build intricate architectures. 

It needs professionals who can ship reliable solutions that create value. 

You can be the person who built something impressive.

Or you can be the person who solved the problem.

Choose reliability and simplicity. Choose solutions that work.

***

Follow Simone Fuscone to help me to create more content
Subscribe to get notified when I post 
Share the post with your network

taplio

The Battle

Alberto Chierici

Simone Fuscone

Followers

Posts

Engagements

Likes

Performance ScoreSimone Fuscone's LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

Alberto Chierici

Simone Fuscone

Followers

Posts

Engagements

Likes

Performance Score.css-goub6s{font-size:var(--chakra-fontSizes-md);font-weight:500;color:var(--chakra-colors-gray-600);display:block;margin-top:var(--chakra-space-1);}Simone Fuscone's LinkedIn Performance

Ready to grow your LinkedIn presence?

Performance ScoreSimone Fuscone's LinkedIn Performance